Merge pull request #180 from ctlearn-project/aitrigger

AI-based trigger system

config/example_config.yml

@@ -11,13 +11,13 @@
 
 # List of reconstruction tasks. For now it's recommended to you single task learning.
 # Valid options:
-#     - 'particletype'
+#     - 'type'
 #     - 'energy'
 #     - 'direction'
 # These options have to be in consonance with the 'event_info' and 
 # 'transforms' in 'Data'.
 # Requirements/recommendations:
-#   - 'particletype':
+#   - 'type':
 #       event_info:
 #           - 'true_shower_primary_id'
 #       transforms:
@@ -66,8 +66,9 @@ Data:
     #   - 'LST_MAGIC_MAGICCam': MAGIC Telescopes
     #   - 'MST_MST_FlashCam': MST Telescope with FlashCAM Camera
     #   - 'MST_MST_NectarCam': MST Telescope with NectarCAM Camera
+    #   - 'SST_1M_DigiCam4': SST1M Telescope with DigiCam Camera
     #   - 'SST_SCT_SCTCam': SCT Telescope and Camera
-    #   - 'SST_ASTRI_ASTRICam': SST-2M ASTRI Telescope and CHECH Camera
+    #   - 'SST_ASTRI_ASTRICam': SST-2M ASTRI Telescope and CHEC Camera
     # Camera images from the following telescopes can be read, but writers
     # for the data are not yet available:
     #   - FACT, HESS-I, HESS-II, VERITAS
@@ -90,11 +91,11 @@ Data:
     # The recommended way of applying quality cuts to the data.
     # Only available for mono analysis at the moment.
     # Optional dict.
-    parameter_selection: 
+    parameter_selection:
         - {col_name: "hillas_intensity", min_value: 50.0}
         - {col_name: "leakage_intensity_width_2", max_value: 0.2}
 
-    # Not the recommended way of applying quality cuts to the data. 
+    # Not the recommended way of applying quality cuts to the data:
     # PyTables cut condition to apply selection cuts to events.
     # Optional string. Default: don't apply any cuts
     # See https://www.pytables.org/usersguide/condition_syntax.html
@@ -151,11 +152,33 @@ Data:
     # Optional integer. Default: use default random initialization.
     seed: 1234
 
-    # Image channels to load.
-    # Optional list of strings. Default: ['image']
-    # Valid options are 'image', 'peak_time' 'cleaned_image' and 'clean_peak_time', or whichever columns are in
-    # the telescope tables of your data files.
-    image_channels: ['image', 'peak_time']
+    # Waveform settings to load.
+    # Optional dicts. Default: None
+    #  - Valid options for 'waveform_type' are 'raw' (R0) or 'calibrated' (R1).
+    #  - 'waveform_sequence_length' is an integer corresponding to the length
+    #    of the sequence around the shower maximum
+    #  - Valid options for 'waveform_format' are 'timechannel_first' or 'timechannel_last'.
+    #  - 'waveform_r0pedsub' is a boolean weather to calculate and subtract the pedestal level
+    #    per pixel outside the sequence aound the shower maximum (only selectable for R0).
+    waveform_settings:
+            waveform_type: 'raw'
+            waveform_sequence_length: 5
+            waveform_format: 'timechannel_last'
+            waveform_r0pedsub: True
+
+    # Image settings to load.
+    # Optional dicts. Default: None
+    #  - Valid options for 'image_channels' are 'image', 'peak_time' 'cleaned_image' and 'clean_peak_time',
+    #    or whichever columns are in the telescope tables of your data files.
+    image_settings:
+        image_channels: ['image', 'peak_time']
+
+    # Parameter settings to load.
+    # Optional dicts. Default: None
+    #  - Valid options for 'parameter_list' are 'image', 'peak_time' 'cleaned_image' and 'clean_peak_time',
+    #    or whichever columns are in the telescope tables of your data files.
+    #parameter_settings:
+    #    parameter_list: ['hillas_intensity', 'leakage_intensity_width_2']
 
     # Settings passed directly as arguments to ImageMapper.
     # Optional dictionary. Default: {}
@@ -193,6 +216,7 @@ Data:
             'LSTCam': 'bilinear_interpolation'
             'FlashCam': 'bilinear_interpolation'
             'NectarCam': 'bilinear_interpolation'
+            'DigiCam': 'bilinear_interpolation'
             'SCTCam': 'oversampling'
             'CHEC': 'oversampling'
             'MAGICCam': 'bilinear_interpolation'
@@ -205,6 +229,7 @@ Data:
             'LSTCam': 2
             'FlashCam': 1
             'NectarCam': 2
+            'DigiCam': 2
             'SCTCam': 0
             'CHEC': 0
             'MAGICCam': 2
@@ -267,10 +292,10 @@ Input:
     # Optional integer. Default: 1
     batch_size_per_worker: 32
 
-    # Whether to concatenate the stereo images from an event to feed it
+    # Whether to stack the stereo images from an event to feed it
     # to a monoscopic model.
     # Optional boolean. Default: false
-    concat_telescopes: false
+    stack_telescope_images: false
 
 # Settings for the TensorFlow model. The options in this and the
 # Model Parameters section are passed to the Estimator model_fn
@@ -442,7 +467,7 @@ Model Parameters:
     # For the classification task, the class names are passed into the dict
     # in order of class label (0 to n-1).
     standard_head:
-        particletype: {class_names: ['proton', 'gamma'], fc_head: [512, 256, 2], weight: 1.0}
+        type: {class_names: ['proton', 'gamma'], fc_head: [512, 256, 2], weight: 1.0}
         energy: {fc_head: [512, 256, 1], weight: 1.0}
         direction: {fc_head: [512, 256, 2], weight: 1.0}
 

ctlearn/build_irf.py

@@ -312,11 +312,8 @@ def main():
             energy_max=u.Quantity(p["mc_header"]["energy_range_max"].max(), u.TeV),
             spectral_index=p["mc_header"]["spectral_index"][0],
             max_impact=u.Quantity(p["mc_header"]["max_scatter_range"].max(), u.m),
-            viewcone=u.Quantity(
-                p["mc_header"]["max_viewcone_radius"][0]
-                - p["mc_header"]["min_viewcone_radius"][0],
-                u.deg,
-            ),
+            viewcone_min=u.Quantity(np.around(p["mc_header"]["min_viewcone_radius"][0], decimals=2), u.deg),
+            viewcone_max=u.Quantity(np.around(p["mc_header"]["max_viewcone_radius"][0], decimals=2), u.deg),
         )
         p["simulation_info"] = simulation_info
         p["simulated_spectrum"] = PowerLaw.from_simulation(simulation_info, T_OBS)

ctlearn/data_loader.py

@@ -13,7 +13,7 @@ def __init__(
         mode="train",
         class_names=None,
         shuffle=True,
-        concat_telescopes=False,
+        stack_telescope_images=False,
     ):
         "Initialization"
         self.DL1DataReaderDL1DH = DL1DataReaderDL1DH
@@ -22,13 +22,14 @@ def __init__(
         self.mode = mode
         self.class_names = class_names
         self.shuffle = shuffle
-        self.concat_telescopes = concat_telescopes
+        self.stack_telescope_images = stack_telescope_images
         self.on_epoch_end()
 
         # Decrypt the example description
         # Features
         self.singleimg_shape = None
         self.trg_pos, self.trg_shape = None, None
+        self.trgpatch_pos, self.trgpatch_shape = None, None
         self.pon_pos = None
         self.pointing = []
         self.wvf_pos, self.wvf_shape = None, None
@@ -42,23 +43,24 @@ def __init__(
         self.mjd_list, self.milli_list, self.nano_list = [], [], []
         # Labels
         self.prt_pos, self.enr_pos, self.drc_pos = None, None, None
-        self.prt_labels, self.enr_labels, self.alt_labels, self.az_labels = (
-            [],
-            [],
-            [],
-            [],
-        )
+        self.prt_labels = []
+        self.enr_labels = []
+        self.alt_labels, self.az_labels = [], []
+        self.trgpatch_labels = []
         self.energy_unit = None
 
         for i, desc in enumerate(self.DL1DataReaderDL1DH.example_description):
-            if "trigger" in desc["name"]:
+            if "HWtrigger" in desc["name"]:
                 self.trg_pos = i
                 self.trg_shape = desc["shape"]
             elif "pointing" in desc["name"]:
                 self.pon_pos = i
             elif "waveform" in desc["name"]:
                 self.wvf_pos = i
                 self.wvf_shape = desc["shape"]
+            elif "trigger_patch" in desc["name"]:
+                self.trgpatch_pos = i
+                self.trgpatch_shape = desc["shape"]
             elif "image" in desc["name"]:
                 self.img_pos = i
                 self.img_shape = desc["shape"]
@@ -91,7 +93,7 @@ def __init__(
                 self.img_shape[3],
             )
         # Reshape inputs into proper dimensions for the stereo analysis with merged models
-        if self.concat_telescopes:
+        if self.stack_telescope_images:
             self.img_shape = (
                 self.img_shape[1],
                 self.img_shape[2],
@@ -131,7 +133,10 @@ def __data_generation(self, batch_indices):
                 energy = np.empty((self.batch_size))
             if self.drc_pos is not None:
                 direction = np.empty((self.batch_size, 2))
-
+            if self.trgpatch_pos is not None:
+                trigger_patches_true_image_sum = np.empty(
+                    (self.batch_size, *self.trgpatch_shape)
+                )
         # Generate data
         for i, index in enumerate(batch_indices):
             event = self.DL1DataReaderDL1DH[index]
@@ -157,15 +162,19 @@ def __data_generation(self, batch_indices):
                     energy[i] = event[self.enr_pos]
                 if self.drc_pos is not None:
                     direction[i] = event[self.drc_pos]
+                if self.trgpatch_pos is not None:
+                    trigger_patches_true_image_sum[i] = event[self.trgpatch_pos]
             else:
                 # Save all labels for the prediction phase
                 if self.prt_pos is not None:
                     self.prt_labels.append(np.float32(event[self.prt_pos]))
                 if self.enr_pos is not None:
-                    self.enr_labels.append(event[self.enr_pos][0])
+                    self.enr_labels.append(np.float32(event[self.enr_pos][0]))
                 if self.drc_pos is not None:
-                    self.alt_labels.append(event[self.drc_pos][0])
-                    self.az_labels.append(event[self.drc_pos][1])
+                    self.alt_labels.append(np.float32(event[self.drc_pos][0]))
+                    self.az_labels.append(np.float32(event[self.drc_pos][1]))
+                if self.trgpatch_pos is not None:
+                    self.trgpatch_labels.append(np.float32(event[self.trgpatch_pos]))
                 # Save pointing
                 if self.pon_pos is not None:
                     self.pointing.append(event[self.pon_pos])
@@ -198,7 +207,7 @@ def __data_generation(self, batch_indices):
         labels = {}
         if self.mode == "train":
             if self.prt_pos is not None:
-                labels["particletype"] = tf.keras.utils.to_categorical(
+                labels["type"] = tf.keras.utils.to_categorical(
                     particletype,
                     num_classes=self.DL1DataReaderDL1DH.num_classes,
                 )
@@ -212,6 +221,9 @@ def __data_generation(self, batch_indices):
             if self.drc_pos is not None:
                 labels["direction"] = direction
                 label = direction
+            if self.trgpatch_pos is not None and self.DL1DataReaderDL1DH.reco_cherenkov_photons:
+                labels["cherenkov_photons"] = trigger_patches_true_image_sum
+                label = trigger_patches_true_image_sum
 
         # Temp fix till keras support class weights for multiple outputs or I wrote custom loss
         # https://github.com/keras-team/keras/issues/11735

ctlearn/default_config_files/CNNRNN.yml

@@ -1,11 +1,13 @@
 Data:
     mode: 'stereo'
-    image_channels: ['image', 'peak_time']
+    image_settings:
+        image_channels: ['image', 'peak_time']
     mapping_settings:
         mapping_method:
             'LSTCam': 'bilinear_interpolation'
             'FlashCam': 'bilinear_interpolation'
             'NectarCam': 'bilinear_interpolation'
+            'DigiCam': 'bilinear_interpolation'
             'CHEC': 'oversampling'
             'SCTCam': 'oversampling'
             'LSTSiPMCam': 'bilinear_interpolation'
@@ -14,17 +16,18 @@ Data:
             'LSTCam': 2
             'FlashCam': 2
             'NectarCam': 2
+            'DigiCam': 2
             'CHEC': 0
             'SCTCam': 0
             'LSTSiPMCam': 2
             'MAGICCam': 2
 Input:
     batch_size_per_worker: 16
-    concat_telescopes: false
+    stack_telescope_images: false
 Model:
     name: 'CNNRNN'
     backbone: {module: 'cnn_rnn', function: 'cnn_rnn_model'}
-    engine_img: {module: 'basic', function: 'conv_block'}
+    image_engine: {module: 'basic', function: 'conv_block'}
     head: {module: 'head', function: 'standard_head'}
 Model Parameters:
     attention: {mechanism: 'Squeeze-and-Excitation', ratio: 16}
@@ -39,11 +42,11 @@ Model Parameters:
             bottleneck: null
             batchnorm: false
     standard_head:
-        particletype: {fc_head: [], weight: 1.0}
+        type: {fc_head: [], weight: 1.0}
         energy: {fc_head: [], weight: 1.0}
         direction: {fc_head: [], weight: 1.0}
 Training:
-    validation_split: 0.05
+    validation_split: 0.1
     num_epochs: 5
     verbose: 2
     workers: 1

ctlearn/default_config_files/SingleCNN.yml

@@ -1,11 +1,13 @@
 Data:
     mode: 'mono'
-    image_channels: ['image', 'peak_time']
+    image_settings:
+        image_channels: ['image', 'peak_time']
     mapping_settings:
         mapping_method:
             'LSTCam': 'bilinear_interpolation'
             'FlashCam': 'bilinear_interpolation'
             'NectarCam': 'bilinear_interpolation'
+            'DigiCam': 'bilinear_interpolation'
             'CHEC': 'oversampling'
             'SCTCam': 'oversampling'
             'LSTSiPMCam': 'bilinear_interpolation'
@@ -14,17 +16,18 @@ Data:
             'LSTCam': 2
             'FlashCam': 2
             'NectarCam': 2
+            'DigiCam': 2
             'CHEC': 0
             'SCTCam': 0
             'LSTSiPMCam': 2
             'MAGICCam': 2
 Input:
     batch_size_per_worker: 64
-    concat_telescopes: false
+    stack_telescope_images: false
 Model:
     name: 'SingleCNN'
     backbone: {module: 'single_cnn', function: 'single_cnn_model'}
-    engine_img: {module: 'basic', function: 'conv_block'}
+    image_engine: {module: 'basic', function: 'conv_block'}
     head: {module: 'head', function: 'standard_head'}
 Model Parameters:
     attention: {mechanism: 'Squeeze-and-Excitation', ratio: 16}
@@ -39,7 +42,7 @@ Model Parameters:
             bottleneck: null
             batchnorm: false
     standard_head:
-        particletype: {fc_head: [512, 256], weight: 1.0}
+        type: {fc_head: [512, 256], weight: 1.0}
         energy: {fc_head: [512, 256], weight: 1.0}
         direction: {fc_head: [512, 256], weight: 1.0}
 Training:

ctlearn/default_config_files/TRN.yml

@@ -1,11 +1,13 @@
 Data:
     mode: 'mono'
-    image_channels: ['image', 'peak_time']
+    image_settings:
+        image_channels: ['image', 'peak_time']
     mapping_settings:
         mapping_method:
             'LSTCam': 'bilinear_interpolation'
             'FlashCam': 'bilinear_interpolation'
             'NectarCam': 'bilinear_interpolation'
+            'DigiCam': 'bilinear_interpolation'
             'CHEC': 'oversampling'
             'SCTCam': 'oversampling'
             'LSTSiPMCam': 'bilinear_interpolation'
@@ -14,17 +16,18 @@ Data:
             'LSTCam': 2
             'FlashCam': 2
             'NectarCam': 2
+            'DigiCam': 2
             'CHEC': 0
             'SCTCam': 0
             'LSTSiPMCam': 2
             'MAGICCam': 2
 Input:
     batch_size_per_worker: 64
-    concat_telescopes: false
+    stack_telescope_images: false
 Model:
     name: 'ThinResNet'
     backbone: {module: 'single_cnn', function: 'single_cnn_model'}
-    engine_img: {module: 'resnet', function: 'stacked_res_blocks'}
+    image_engine: {module: 'resnet', function: 'stacked_res_blocks'}
     head: {module: 'head', function: 'standard_head'}
 Model Parameters:
     attention: {mechanism: 'Squeeze-and-Excitation', ratio: 16}
@@ -37,7 +40,7 @@ Model Parameters:
                 - {filters: 128, blocks: 3}
                 - {filters: 256, blocks: 3}
     standard_head:
-        particletype: {fc_head: [512, 256], weight: 1.0}
+        type: {fc_head: [512, 256], weight: 1.0}
         energy: {fc_head: [512, 256], weight: 1.0}
         direction: {fc_head: [512, 256], weight: 1.0}
 Training: