initiaö commit

7 files changed, 567 insertions, 0 deletions

diff --git a/cfg/adept/adept-level1-run1.json b/cfg/adept/adept-level1-run1.json
new file mode 100644
index 0000000..22f6abb
--- /dev/null
+++ b/cfg/adept/adept-level1-run1.json
@@ -0,0 +1,81 @@
+{
+   "model_path": "adept_level1",
+   "datatype": "adept",
+   "dataset": "ADEPT",
+   "learning_rate": 0.0001,
+   "num_updates": 0,
+   "max_epochs": 1000,
+   "max_updates": 400000,
+   "phases": {
+      "start_inner_loop": 60000,
+      "shufleslots_end": 30000,
+      "entity_pretraining_phase2_end": 50000,
+      "entity_pretraining_phase1_end": 30000,
+      "background_pretraining_end": 0
+   },
+   "defaults": {
+      "num_workers": 2,
+      "prefetch_factor": 2,
+      "statistics_offset": 10,
+      "load_optimizers": false,
+      "teacher_forcing": 10,
+      "skip_frames": 3,
+      "error_dropout": 0.1
+   },
+   "bptt": {
+      "bptt_start_timestep": 0,
+      "bptt_steps": 1,
+      "bptt_steps_max": 2,
+      "increase_bptt_steps_every": 200000
+   },
+   "model": {
+    "level": 1,
+    "batch_size": 16,
+    "num_objects": 7,
+    "img_channels": 3,
+    "input_size": [
+     320,
+     480
+    ],
+    "latent_size": [
+     20,
+     30
+    ],
+    "gestalt_size": 96,
+    "bottleneck": "binar",
+    "position_regularizer": 0.01,
+    "time_regularizer": 0.1,
+    "encoder_regularizer": 0.333333,
+    "inner_loop_enabled": false,
+    "encoder": {
+     "channels": 48,
+     "level1_channels": 24,
+     "num_layers": 3,
+     "reg_lambda": 1e-10
+    },
+    "predictor": {
+     "heads": 2,
+     "layers": 2,
+     "channels_multiplier": 2,
+     "reg_lambda": 1e-10,
+     "transformer_type": "standard"
+    },
+    "decoder": {
+     "channels": 48,
+     "level1_channels": 24,
+     "num_layers": 5
+    },
+    "background": {
+     "learning_rate": 0.0001,
+     "learning_rate_old": 0.0001,
+     "num_layers": 1,
+     "latent_channels": 48,
+     "level1_channels": 24,
+     "gestalt_size": 8,
+     "flow": false
+    },
+    "update_module": {
+     "reg_lambda": 5e-06
+    }
+   }
+  }
\ No newline at end of file
diff --git a/cfg/adept/adept-level1-run2.json b/cfg/adept/adept-level1-run2.json
new file mode 100644
index 0000000..7263eff
--- /dev/null
+++ b/cfg/adept/adept-level1-run2.json
@@ -0,0 +1,81 @@
+{
+   "model_path": "adept_level1",
+   "datatype": "adept",
+   "dataset": "ADEPT",
+   "learning_rate": 3.333e-05,
+   "num_updates": 400000,
+   "max_epochs": 1000,
+   "max_updates": 600000,
+   "phases": {
+      "start_inner_loop": 60000,
+      "shufleslots_end": 30000,
+      "entity_pretraining_phase2_end": 50000,
+      "entity_pretraining_phase1_end": 30000,
+      "background_pretraining_end": 0
+   },
+   "defaults": {
+      "num_workers": 2,
+      "prefetch_factor": 2,
+      "statistics_offset": 10,
+      "load_optimizers": false,
+      "teacher_forcing": 10,
+      "skip_frames": 3,
+      "error_dropout": 0.1
+   },
+   "bptt": {
+      "bptt_start_timestep": 0,
+      "bptt_steps": 2,
+      "bptt_steps_max": 2,
+      "increase_bptt_steps_every": 200000
+   },
+   "model": {
+    "level": 1,
+    "batch_size": 16,
+    "num_objects": 7,
+    "img_channels": 3,
+    "input_size": [
+     320,
+     480
+    ],
+    "latent_size": [
+     20,
+     30
+    ],
+    "gestalt_size": 96,
+    "bottleneck": "binar",
+    "position_regularizer": 0.01,
+    "time_regularizer": 0.1,
+    "encoder_regularizer": 0.333333,
+    "inner_loop_enabled": true,
+    "encoder": {
+     "channels": 48,
+     "level1_channels": 24,
+     "num_layers": 3,
+     "reg_lambda": 1e-10
+    },
+    "predictor": {
+     "heads": 2,
+     "layers": 2,
+     "channels_multiplier": 2,
+     "reg_lambda": 1e-10,
+     "transformer_type": "standard"
+    },
+    "decoder": {
+     "channels": 48,
+     "level1_channels": 24,
+     "num_layers": 5
+    },
+    "background": {
+     "learning_rate": 0.0001,
+     "learning_rate_old": 0.0001,
+     "num_layers": 1,
+     "latent_channels": 48,
+     "level1_channels": 24,
+     "gestalt_size": 8,
+     "flow": false
+    },
+    "update_module": {
+     "reg_lambda": 5e-06
+    }
+   }
+  }
\ No newline at end of file
diff --git a/cfg/adept/adept-level1-run3.json b/cfg/adept/adept-level1-run3.json
new file mode 100644
index 0000000..c715e97
--- /dev/null
+++ b/cfg/adept/adept-level1-run3.json
@@ -0,0 +1,81 @@
+{
+   "model_path": "adept_level1",
+   "datatype": "adept",
+   "dataset": "ADEPT",
+   "learning_rate": 3.333e-05,
+   "num_updates": 600000,
+   "max_epochs": 1000,
+   "max_updates": 800000,
+   "phases": {
+      "start_inner_loop": 60000,
+      "shufleslots_end": 30000,
+      "entity_pretraining_phase2_end": 50000,
+      "entity_pretraining_phase1_end": 30000,
+      "background_pretraining_end": 0
+   },
+   "defaults": {
+      "num_workers": 2,
+      "prefetch_factor": 2,
+      "statistics_offset": 10,
+      "load_optimizers": false,
+      "teacher_forcing": 10,
+      "skip_frames": 3,
+      "error_dropout": 0.1
+   },
+   "bptt": {
+      "bptt_start_timestep": 0,
+      "bptt_steps": 2,
+      "bptt_steps_max": 2,
+      "increase_bptt_steps_every": 200000
+   },
+   "model": {
+    "level": 1,
+    "batch_size": 16,
+    "num_objects": 7,
+    "img_channels": 3,
+    "input_size": [
+     320,
+     480
+    ],
+    "latent_size": [
+     20,
+     30
+    ],
+    "gestalt_size": 96,
+    "bottleneck": "binar",
+    "position_regularizer": 0.01,
+    "time_regularizer": 0.1,
+    "encoder_regularizer": 0.333333,
+    "inner_loop_enabled": true,
+    "encoder": {
+     "channels": 48,
+     "level1_channels": 24,
+     "num_layers": 3,
+     "reg_lambda": 1e-10
+    },
+    "predictor": {
+     "heads": 2,
+     "layers": 2,
+     "channels_multiplier": 2,
+     "reg_lambda": 1e-10,
+     "transformer_type": "standard"
+    },
+    "decoder": {
+     "channels": 48,
+     "level1_channels": 24,
+     "num_layers": 5
+    },
+    "background": {
+     "learning_rate": 0.0001,
+     "learning_rate_old": 0.0001,
+     "num_layers": 1,
+     "latent_channels": 48,
+     "level1_channels": 24,
+     "gestalt_size": 8,
+     "flow": false
+    },
+    "update_module": {
+     "reg_lambda": 5e-06
+    }
+   }
+  }
\ No newline at end of file
diff --git a/cfg/adept/adept-level1-run4.json b/cfg/adept/adept-level1-run4.json
new file mode 100644
index 0000000..c750f0d
--- /dev/null
+++ b/cfg/adept/adept-level1-run4.json
@@ -0,0 +1,81 @@
+{
+   "model_path": "adept_level2",
+   "datatype": "adept",
+   "dataset": "ADEPT",
+   "learning_rate": 3.333e-05,
+   "num_updates": 800000,
+   "max_epochs": 1000,
+   "max_updates": 1150000,
+   "phases": {
+      "start_inner_loop": 60000,
+      "shufleslots_end": 30000,
+      "entity_pretraining_phase2_end": 50000,
+      "entity_pretraining_phase1_end": 30000,
+      "background_pretraining_end": 0
+   },
+   "defaults": {
+      "num_workers": 2,
+      "prefetch_factor": 2,
+      "statistics_offset": 10,
+      "load_optimizers": false,
+      "teacher_forcing": 10,
+      "skip_frames": 3,
+      "error_dropout": 0.1
+   },
+   "bptt": {
+      "bptt_start_timestep": 0,
+      "bptt_steps": 2,
+      "bptt_steps_max": 2,
+      "increase_bptt_steps_every": 200000
+   },
+   "model": {
+    "level": 2,
+    "batch_size": 16,
+    "num_objects": 7,
+    "img_channels": 3,
+    "input_size": [
+     320,
+     480
+    ],
+    "latent_size": [
+     20,
+     30
+    ],
+    "gestalt_size": 96,
+    "bottleneck": "binar",
+    "position_regularizer": 0.01,
+    "time_regularizer": 0.1,
+    "encoder_regularizer": 0.333333,
+    "inner_loop_enabled": true,
+    "encoder": {
+     "channels": 48,
+     "level1_channels": 24,
+     "num_layers": 3,
+     "reg_lambda": 1e-10
+    },
+    "predictor": {
+     "heads": 2,
+     "layers": 2,
+     "channels_multiplier": 2,
+     "reg_lambda": 1e-10,
+     "transformer_type": "standard"
+    },
+    "decoder": {
+     "channels": 48,
+     "level1_channels": 24,
+     "num_layers": 5
+    },
+    "background": {
+     "learning_rate": 0.0001,
+     "learning_rate_old": 0.0001,
+     "num_layers": 1,
+     "latent_channels": 48,
+     "level1_channels": 24,
+     "gestalt_size": 8,
+     "flow": false
+    },
+    "update_module": {
+     "reg_lambda": 5e-06
+    }
+   }
+  }
\ No newline at end of file
diff --git a/cfg/clevrer/clevrer-level1-run1.json b/cfg/clevrer/clevrer-level1-run1.json
new file mode 100644
index 0000000..da64d01
--- /dev/null
+++ b/cfg/clevrer/clevrer-level1-run1.json
@@ -0,0 +1,81 @@
+{
+  "model_path": "clevrer_level1",
+  "datatype": "clevrer",
+  "dataset": "CLEVRER",
+  "learning_rate": 0.0001,
+  "num_updates": 0,
+  "max_epochs": 1000,
+  "max_updates": 200000,
+  "phases": {
+     "start_inner_loop": 100000,
+     "shufleslots_end": 30000,
+     "entity_pretraining_phase2_end": 50000,
+     "entity_pretraining_phase1_end": 30000,
+     "background_pretraining_end": 0
+  },
+  "defaults": {
+     "num_workers": 2,
+     "prefetch_factor": 2,
+     "statistics_offset": 10,
+     "load_optimizers": false,
+     "teacher_forcing": 10,
+     "skip_frames": 2,
+     "error_dropout": 0.1
+  },
+  "bptt": {
+     "bptt_start_timestep": 0,
+     "bptt_steps": 1,
+     "bptt_steps_max": 2,
+     "increase_bptt_steps_every": 200000
+  },
+  "model": {
+   "level": 1,
+   "batch_size": 32,
+   "num_objects": 6,
+   "img_channels": 3,
+   "input_size": [
+    320,
+    480
+   ],
+   "latent_size": [
+    20,
+    30
+   ],
+   "gestalt_size": 96,
+   "bottleneck": "binar",
+   "position_regularizer": 0.01,
+   "time_regularizer": 0.1,
+   "encoder_regularizer": 0.333333,
+   "inner_loop_enabled": false,
+   "encoder": {
+    "channels": 48,
+    "level1_channels": 24,
+    "num_layers": 3,
+    "reg_lambda": 1e-10
+   },
+   "predictor": {
+    "heads": 2,
+    "layers": 2,
+    "channels_multiplier": 2,
+    "reg_lambda": 1e-10,
+    "transformer_type": "shared"
+   },
+   "decoder": {
+    "channels": 48,
+    "level1_channels": 24,
+    "num_layers": 3
+   },
+   "background": {
+    "learning_rate": 0.0001,
+    "learning_rate_old": 0.0001,
+    "num_layers": 1,
+    "latent_channels": 48,
+    "level1_channels": 24,
+    "gestalt_size": 8,
+    "flow": false
+   },
+   "update_module": {
+    "reg_lambda": 5e-08
+   }
+  }
+ }
\ No newline at end of file
diff --git a/cfg/clevrer/clevrer-level1-run2.json b/cfg/clevrer/clevrer-level1-run2.json
new file mode 100644
index 0000000..0d2d23e
--- /dev/null
+++ b/cfg/clevrer/clevrer-level1-run2.json
@@ -0,0 +1,81 @@
+{
+  "model_path": "clevrer_level1",
+  "datatype": "clevrer",
+  "dataset": "CLEVRER",
+  "learning_rate": 3.3333e-05,
+  "num_updates": 200000,
+  "max_epochs": 1000,
+  "max_updates": 600000,
+  "phases": {
+     "start_inner_loop": 100000,
+     "shufleslots_end": 30000,
+     "entity_pretraining_phase2_end": 50000,
+     "entity_pretraining_phase1_end": 30000,
+     "background_pretraining_end": 0
+  },
+  "defaults": {
+     "num_workers": 2,
+     "prefetch_factor": 2,
+     "statistics_offset": 10,
+     "load_optimizers": false,
+     "teacher_forcing": 10,
+     "skip_frames": 2,
+     "error_dropout": 0.1
+  },
+  "bptt": {
+     "bptt_start_timestep": 0,
+     "bptt_steps": 3,
+     "bptt_steps_max": 3,
+     "increase_bptt_steps_every": 200000
+  },
+  "model": {
+   "level": 1,
+   "batch_size": 32,
+   "num_objects": 6,
+   "img_channels": 3,
+   "input_size": [
+    320,
+    480
+   ],
+   "latent_size": [
+    20,
+    30
+   ],
+   "gestalt_size": 96,
+   "bottleneck": "binar",
+   "position_regularizer": 0.01,
+   "time_regularizer": 0.1,
+   "encoder_regularizer": 0.333333,
+   "inner_loop_enabled": true,
+   "encoder": {
+    "channels": 48,
+    "level1_channels": 24,
+    "num_layers": 3,
+    "reg_lambda": 1e-10
+   },
+   "predictor": {
+    "heads": 2,
+    "layers": 2,
+    "channels_multiplier": 2,
+    "reg_lambda": 1e-10,
+    "transformer_type": "shared"
+   },
+   "decoder": {
+    "channels": 48,
+    "level1_channels": 24,
+    "num_layers": 3
+   },
+   "background": {
+    "learning_rate": 0.0001,
+    "learning_rate_old": 0.0001,
+    "num_layers": 1,
+    "latent_channels": 48,
+    "level1_channels": 24,
+    "gestalt_size": 8,
+    "flow": false
+   },
+   "update_module": {
+    "reg_lambda": 5e-08  
+   }
+  }
+ }
\ No newline at end of file
diff --git a/cfg/clevrer/clevrer-level1-run3.json b/cfg/clevrer/clevrer-level1-run3.json
new file mode 100644
index 0000000..df31316
--- /dev/null
+++ b/cfg/clevrer/clevrer-level1-run3.json
@@ -0,0 +1,81 @@
+{
+  "model_path": "clevrer_level1",
+  "datatype": "clevrer",
+  "dataset": "CLEVRER",
+  "learning_rate": 3.3333e-05,
+  "num_updates": 600000,
+  "max_epochs": 1000,
+  "max_updates": 800000,
+  "phases": {
+     "start_inner_loop": 100000,
+     "shufleslots_end": 30000,
+     "entity_pretraining_phase2_end": 50000,
+     "entity_pretraining_phase1_end": 30000,
+     "background_pretraining_end": 0
+  },
+  "defaults": {
+     "num_workers": 2,
+     "prefetch_factor": 2,
+     "statistics_offset": 10,
+     "load_optimizers": false,
+     "teacher_forcing": 10,
+     "skip_frames": 2,
+     "error_dropout": 0.1
+  },
+  "bptt": {
+     "bptt_start_timestep": 0,
+     "bptt_steps": 4,
+     "bptt_steps_max": 4,
+     "increase_bptt_steps_every": 200000
+  },
+  "model": {
+   "level": 1,
+   "batch_size": 32,
+   "num_objects": 6,
+   "img_channels": 3,
+   "input_size": [
+    320,
+    480
+   ],
+   "latent_size": [
+    20,
+    30
+   ],
+   "gestalt_size": 96,
+   "bottleneck": "binar",
+   "position_regularizer": 0.01,
+   "time_regularizer": 0.1,
+   "encoder_regularizer": 0.333333,
+   "inner_loop_enabled": true,
+   "encoder": {
+    "channels": 48,
+    "level1_channels": 24,
+    "num_layers": 3,
+    "reg_lambda": 1e-10
+   },
+   "predictor": {
+    "heads": 2,
+    "layers": 2,
+    "channels_multiplier": 2,
+    "reg_lambda": 1e-10,
+    "transformer_type": "shared"
+   },
+   "decoder": {
+    "channels": 48,
+    "level1_channels": 24,
+    "num_layers": 3
+   },
+   "background": {
+    "learning_rate": 0.0001,
+    "learning_rate_old": 0.0001,
+    "num_layers": 1,
+    "latent_channels": 48,
+    "level1_channels": 24,
+    "gestalt_size": 8,
+    "flow": false
+   },
+   "update_module": {
+    "reg_lambda": 5e-08  
+   }
+  }
+ }
\ No newline at end of file