Merge pull request #107 from openspyrit/train

Tutorial to train a reconstruction network: Network can be selected

spyrit/tutorial/train.py

@@ -47,6 +47,7 @@
     parser.add_argument("--stat_root",  type=str,   default="./stat/", help="Path to precomputed data")
     parser.add_argument("--arch",       type=str,   default="dc-net", help="Choose among 'dc-net','pinv-net',")
     parser.add_argument("--denoi",      type=str,   default="unet", help="Choose among 'cnn','cnnbn', 'unet'")
+    parser.add_argument("--device",     type=str,   default="", help="Choose among 'cuda','cpu'")
     #parser.add_argument("--no_denoi",   default=False, action='store_true', help="No denoising layer")
 
 
@@ -68,19 +69,17 @@
     opt.model_root = Path(opt.model_root)
     opt.data_root = Path(opt.data_root)
 
-    #if opt.data == 'stl10':
-    #    opt.data_root = '../../../data/'
-
-    opt.arch = "upgd"
-
     print(opt)
 
     #==========================================================================
     # 0. Setting up parameters for training
     #==========================================================================
     # The device of the machine, number of workers...
     # 
-    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+    if opt.device: 
+        device = torch.device(opt.device)
+    else:
+        device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
     print(f'Device: {device}')
 
     #==========================================================================

spyrit/tutorial/tuto_core_train.py

@@ -0,0 +1,131 @@
+r"""
+04. Tutorial 2D - Tutorial to train a reconstruction network 
+======================
+This tutorial shows how to train a reconstruction network for 2D single-pixel imaging 
+on stl10. Training is performed by a call to *train.py*. Several parameters allow 
+to modify acquisition, network and training (network architecture), 
+optimisation and the use of tensorboard. 
+ 
+Currently you can train the following networks by modifying the network architecture variable *arch*: 
+ 
+    - 'dc-net': Denoised Completion Network (DCNet). 
+    - 'pinv-net': Pseudo Inverse Network (PinvNet).
+    - 'upgd': Unrolled proximal gradient descent (UPGD). 
+    
+and the denoising variable *denoi*: E
+    - 'cnn': CNN no batch normalization
+    - 'cnnbn': CNN with batch normalization
+    - 'unet': UNet (0.5 M trainable parameters) 
+
+
+[Colab version]: https://colab.research.google.com/github/openspyrit/spyrit/blob/master/spyrit/tutorial/tuto_train_colab.ipynb
+
+"""
+
+###############################################################################
+# Import packages
+import os
+import datetime
+import subprocess
+
+###############################################################################
+# Download covariance matrix. Alternatively install *openspyrit/spas* package:
+# 
+#     spyrit
+#     ├───stat
+#     │   ├───Average_64x64.npy
+#     │   ├───Cov_64x64.npy
+#     ├───spirit
+# 
+
+download_cov = True
+if (download_cov is True):
+    # pip install girder-client
+    import girder_client
+
+    # api Rest url of the warehouse
+    url='https://pilot-warehouse.creatis.insa-lyon.fr/api/v1'
+
+    # Generate the warehouse client
+    gc = girder_client.GirderClient(apiUrl=url)
+
+    # Download the covariance matrix and mean image
+    data_folder = './stat/'
+    dataId_list = [
+            '63935b624d15dd536f0484a5', # for reconstruction (imageNet, 64)
+            '63935a224d15dd536f048496', # for reconstruction (imageNet, 64)
+            ]
+    for dataId in dataId_list:
+        myfile = gc.getFile(dataId)
+        gc.downloadFile(dataId, data_folder + myfile['name'])
+
+    print(f'Created {data_folder}') 
+    os.listdir(data_folder)
+
+###############################################################################
+r""" Train
+You can choose the following parameters:
+ 
+    - Acquisition: 
+        - --img_size: Height / width dimension, default=64
+        - --M: Number of undersampling patterns, default=512
+        - --subs: Among 'var','rect', default="var"
+     
+    - Network and training: 
+        - --data: stl10 or imagenet, default="stl10"
+        - --model_root: Path to model saving files, default='./model/'
+        - --data_root: Path to the dataset, default="./data/"
+    
+        - --N0: Mean maximum total number of photons, default=10
+        - --stat_root: Path to precomputed data, default="./stat/"
+        - --arch: Choose among 'dc-net','pinv-net', 'upgd', default="dc-net"
+        - --denoi: Choose among 'cnn','cnnbn', 'unet', default="unet"
+        - --device", Choose among 'cuda','cpu', default="" (cuda if available)
+
+    - Optimisation:
+        - --num_epochs: Number of training epochs, default=30
+        - --batch_size: Size of each training batch, default=512
+        - --reg: Regularisation Parameter, default=1e-7
+        - --step_size: Scheduler Step Size, default=10
+        - --gamma: Scheduler Decrease Rate, default=0.5
+        - --checkpoint_model: Optional path to checkpoint model, default=""
+        - --checkpoint_interval: Interval between saving model checkpoints, default=0
+        - Training is done with *Adam* optimizer, *MSELoss*
+ 
+    - Tensorboard:
+        - --tb_path: Relative path for Tensorboard experiment tracking logs, default=False
+        - --tb_prof: Code profiler with Tensorboard, default=False
+        - Logging of scalars *train_loss*, *val_loss* and images (dataset example ground-truth and predictions at different epochs).
+"""
+
+###############################################################################
+# In this tutorial, data is perturbed by Poisson noise (100 mean photons) 
+# and undersampling factor of 4, on stl10 dataset.
+# Training is done with default parameters (see above) and using experiment tracking with tensorboard. 
+
+# Parameters
+N0 = 100
+M = 1024
+data_root = './data/'
+data = 'stl10'
+stat_root = './stat'
+
+# Tensorboard logs path
+now = datetime.datetime.now().strftime('%Y-%m-%d_%H-%M')
+tb_path = f'runs/runs_stdl10_n100_m1024/{now}' 
+tb_prof = True # False
+
+# Run train.py
+cmd = ['python', 'spyrit/tutorial/train.py', '--N0', str(N0), '--M', str(M), '--data_root', str(data_root), 
+       '--data', str(data), '--stat_root', str(stat_root), '--tb_path', str(tb_path), '--tb_prof', str(tb_prof), 
+       '--device', str('cpu')]
+subprocess.run(cmd, check=True)
+
+###############################################################################
+# Tensorboard
+#
+# To launch tensorboard, run in a terminal:  
+# tensorboard --logdir <tb_path>
+#
+# Select *SCALARS* or *IMAGES*. More options are available in the top-right corner.
+