Sleepnet (#13)

* Include sleepnet and refactoring

.gitignore

@@ -17,4 +17,6 @@ outputs
 
 src/asleep/torch_hub_cache
 assets/*.joblib.*
-*.ipynb_checkpoints
+*.ipynb_checkpoints
+*.ipynb
+src/asleep/models

MANIFEST.in

@@ -1,3 +1,6 @@
 include tox.ini
 recursive-include tests *.py
-recursive-include assets *.jpg
+recursive-include assets *.jpg *.lzma
+recursive-include src *.yaml
+exclude src/asleep/models *.mdl
+recursive-exclude src/asleep/torch_hub_cache *

README.md

@@ -24,6 +24,8 @@ $ pip install asleep
 ```
 
 # Usage
+All the processing will be much faster after the first time because the model weights will to have to be downloaded
+the first time that the package is used.
 ```shell
 # Process an AX3 file
 $ get_sleep sample.cwa
@@ -88,5 +90,5 @@ TBD
 
 # Acknowledgements
 We would like to thank all our code contributors, manuscript co - authors, and research participants for their help in making this work possible. The
-data processing pipeline of this repository is based on the[step_count](https: // github.com / OxWearables / stepcount  # processing-csv-files) package from our group. Special
-thanks to @ chanshing for his help in developing the package.
+data processing pipeline of this repository is based on the [step_count](https://github.com/OxWearables/stepcount) package from our group. Special
+thanks to @chanshing for his help in developing the package.

hubconf.py

@@ -0,0 +1,26 @@
+import os
+from src.asleep.models import CNNLSTM, weight_init
+import torch
+
+dependencies = ["torch"]
+
+
+def sleepnet(pretrained=True, my_device="cpu", class_num=2, lstm_nn_size=128,
+             dropout_p=0.5, bi_lstm=True, lstm_layer=1):
+    model = CNNLSTM(
+        num_classes=class_num,
+        model_device=my_device,
+        lstm_nn_size=lstm_nn_size,
+        dropout_p=dropout_p,
+        bidrectional=bi_lstm,
+        lstm_layer=lstm_layer,
+    )
+    weight_init(model)
+
+    if pretrained:
+        checkpoint = 'https://download.pytorch.org/models/resnet18-5c106cde.pth'
+        model.load_state_dict(torch.hub.load_state_dict_from_url(checkpoint,
+                                                                 progress=True,
+                                                                 map_location=torch.device(my_device)))
+    model.to(my_device, dtype=torch.float)
+    return model

pyproject.toml

@@ -114,7 +114,8 @@ dependencies = [ # Optional
   "torch",
   "torchvision",
   "transforms3d",
-  "stepcount"
+  "stepcount",
+  "hydra-core",
 ]
 
 # List additional groups of dependencies here (e.g. development

src/asleep/conf/config.yaml

@@ -0,0 +1,18 @@
+defaults:
+  - model: baseline
+  - data: vm_data
+  - _self_
+
+gpu: 0
+num_epoch: 200
+multi_gpu: false
+gpu_ids: [0, 1, 2, 3]  # if more than one gpu
+verbose: false
+patience: 10
+deployment: false
+num_split: 5
+test_size: 0.2
+validation_fold: -1 # validation fold < num_split used for cnn cv
+isDebug: false
+train_master: false
+specific_fold: -1

src/asleep/conf/config_eval.yaml

@@ -0,0 +1,15 @@
+defaults:
+  - model: cnn_lstm_eval
+  - data: deployment
+  - _self_
+
+gpu: -1
+num_epoch: 200
+multi_gpu: false
+gpu_ids: [0, 1, 2, 3]  # if more than one gpu
+verbose: false
+patience: 10
+deployment: true
+num_split: 5
+test_size: 0.2
+validation_fold: -1 # validation fold < num_split used for cnn cv

src/asleep/conf/data/deployment.yaml

@@ -0,0 +1,16 @@
+data_root: /Users/hangy/data
+X_zip_path: "${data.data_root}/X.npy.gz"
+X_path: "${data.data_root}/X.npy"
+PID_path: "${data.data_root}/npid.npy"
+subject_file: ""
+y_pred_path: "${data.data_root}/y_pred.npy"
+y_prob_path: "${data.data_root}/pred_prob.npy"
+pid_pred_path: "${data.data_root}/PID_pred.npy"
+use_gen2: False
+context_path: "${data.data_root}/context.pkl"
+log_path: /home/cxx579/raine/logs
+win_length: 1
+meta_data_size: 0
+test_ratio: 0.2
+val_ratio: 0.2
+num_classes: 5

src/asleep/conf/model/baseline.yaml

@@ -0,0 +1,3 @@
+learning_rate: 0.0001
+isSmall: false
+name: "baseline_WinLen${data.win_length}_BiDi${data.isBidirectional}_Meta${data.meta_data_size}_Aug${augment}"

src/asleep/conf/model/cnn_lstm_eval.yaml

@@ -0,0 +1,12 @@
+deployment: false
+deployment_path: "/home/cxx579/raine/predictions"
+# eval_weight_path: "/data/UKBB/final_models/sleepnet_12_16.mdl"
+isSmall: false
+bi_lstm: True
+lstm_nn_size: 1024
+lstm_layer: 2
+batch_size: 2  # indicate the number of subjects for cnnlstm
+model_name: 'cnn_lstm'
+dropout_p: 0
+multi_task: false
+augment: 'cnn_lstm'