Fake pull request

dataset_multiMNIST/ML_parameters.json

@@ -0,0 +1,112 @@
+{
+
+ "neptune_project" : "dalessioluca/MNIST",
+
+ "simulation" : {
+   "__comment" : "there are 3 types of runs: scratch, resume, pretrained",
+   "type" : "scratch",
+   "MAX_EPOCHS" : 1501,
+   "TEST_FREQUENCY" : 50,
+   "CHECKPOINT_FREQUENCY" : 50,
+   "batch_size" : 128
+ },
+
+ "architecture" : {
+   "__comment" : "architecture parameters, level_zwhere_output is between 0 and n_max_pool included",
+   "dim_zinstance" : 20,
+   "dim_zwhere" : 4,
+   "dim_zbg" : 20,
+   "dim_logit" : 1,
+   "cropped_size" : 28,
+   "n_max_pool" : 4,
+   "level_zwhere_and_logit_output" : 2,
+   "level_background_output" : 4,
+   "n_ch_output_features" : 32,
+   "n_ch_after_first_two_conv" : 32,
+   "concatenate_raw_image_to_fmap" : true
+ },
+
+  "DPP" : {
+    "__comment" : "parameters for the DPP prior",
+    "n_kernels" : 1
+  },
+
+ "input_image" : {
+   "__comment" : "parameters describing the input images",
+   "n_objects_max" : 10,
+   "size_object_min" : 15,
+   "size_object_max" : 50,
+   "length_scale_similarity" : 10.0,
+   "size_raw_image" : 80,
+   "ch_in" : 1
+ },
+
+ "nms" : {
+   "__comment" : "parameters for the non-max-suppression",
+   "overlap_threshold_train" : 0.3,
+   "overlap_threshold_test" : 0.5
+ },
+
+ "GECO_loss" : {
+   "__comment" : "if active=false use ELBO, else use GECO with Log-Likelihood threshold = n_pixels * n_channel * threshold",
+   "is_active" : true,
+
+   "geco_mse_range" : [0.1, 0.8, 0.9],
+   "geco_fgfraction_range" : [0.1, 1.0, 10.0],
+   "geco_ncell_range" : [0.1, 1.0, 10.0],
+
+   "target_fgfraction" : [0.05, 0.15],
+   "target_mse" : [0.25, 0.75],
+   "target_ncell" : [3.0, 5.0],
+
+   "bg_std" : 0.1,
+   "fg_std" : 0.1
+ },
+
+ "optimizer" : {
+   "__comment" : "which optimizer to use",
+   "type" : "adam",
+   "base_lr" : 0.001,
+   "betas" : [0.9, 0.999],
+   "base_lr_similarity" : 0.01,
+   "betas_similarity" : [0.9, 0.999],
+   "base_lr_geco" : 0.001,
+   "betas_geco" : [0.9, 0.999],
+   "weight_decay" : 0.0,
+   "eps" : 1E-8,
+   "scheduler_is_active" : true,
+   "scheduler_type" : "step_LR",
+   "scheduler_step_size" : 500,
+   "scheduler_gamma" : 0.75
+ },
+
+ "shortcut_prob_corr_factor" : {
+  "__comment" : "parameters for the shortcut for porb_corr_factor",
+  "values" : [0.4, 0.0],
+  "times" : [50, 150]
+ },
+
+ "soft_constraint" : {
+
+  "__comment": "all the parameters about the soft constraints",
+
+  "overlap": {
+    "__comment": "cost which discourages masks from overlapping",
+    "strength": 0,
+    "exponent": 1
+  },
+
+  "mask_volume_absolute" : {
+   "__comment" : "cost which discourage masks which are too large or too small",
+   "lower_bound_value": 40,
+   "lower_bound_width": 5,
+   "lower_bound_strength": 0,
+   "lower_bound_exponent": 2,
+   "upper_bound_value": 300,
+   "upper_bound_width": 5,
+   "upper_bound_strength": 0,
+   "upper_bound_exponent": 2
+  }
+ }
+}
+

dataset_multiMNIST/ML_parameters_NO_GRID.json

@@ -0,0 +1,112 @@
+{
+
+ "neptune_project" : "dalessioluca/MNIST",
+
+ "simulation" : {
+   "__comment" : "there are 3 types of runs: scratch, resume, pretrained",
+   "type" : "scratch",
+   "MAX_EPOCHS" : 501,
+   "TEST_FREQUENCY" : 50,
+   "CHECKPOINT_FREQUENCY" : 50,
+   "batch_size" : 128
+ },
+
+ "architecture" : {
+   "__comment" : "architecture parameters, level_zwhere_output is between 0 and n_max_pool included",
+   "dim_zinstance" : 20,
+   "dim_zwhere" : 4,
+   "dim_zbg" : 4,
+   "dim_logit" : 1,
+   "cropped_size" : 28,
+   "n_max_pool" : 4,
+   "level_zwhere_and_logit_output" : 2,
+   "level_background_output" : 4,
+   "n_ch_output_features" : 32,
+   "n_ch_after_first_two_conv" : 32,
+   "concatenate_raw_image_to_fmap" : true
+ },
+
+  "DPP" : {
+    "__comment" : "parameters for the DPP prior",
+    "n_kernels" : 1
+  },
+
+ "input_image" : {
+   "__comment" : "parameters describing the input images",
+   "n_objects_max" : 10,
+   "size_object_min" : 15,
+   "size_object_max" : 50,
+   "length_scale_similarity" : 10.0,
+   "size_raw_image" : 80,
+   "ch_in" : 1
+ },
+
+ "nms" : {
+   "__comment" : "parameters for the non-max-suppression",
+   "overlap_threshold_train" : 0.3,
+   "overlap_threshold_test" : 0.5
+ },
+
+ "GECO_loss" : {
+   "__comment" : "if active=false use ELBO, else use GECO with Log-Likelihood threshold = n_pixels * n_channel * threshold",
+   "is_active" : true,
+
+   "geco_mse_range" : [0.1, 0.8, 0.9],
+   "geco_fgfraction_range" : [0.1, 1.0, 3.0],
+   "geco_ncell_range" : [0.1, 1.0, 3.0],
+
+   "target_fgfraction" : [0.05, 0.25],
+   "target_mse" : [0.25, 0.75],
+   "target_ncell" : [1.2, 7.0],
+
+   "bg_std" : 0.1,
+   "fg_std" : 0.1
+ },
+
+ "optimizer" : {
+   "__comment" : "which optimizer to use",
+   "type" : "adam",
+   "base_lr" : 0.001,
+   "betas" : [0.9, 0.999],
+   "base_lr_similarity" : 0.01,
+   "betas_similarity" : [0.9, 0.999],
+   "base_lr_geco" : 0.001,
+   "betas_geco" : [0.9, 0.999],
+   "weight_decay" : 0.0,
+   "eps" : 1E-8,
+   "scheduler_is_active" : true,
+   "scheduler_type" : "step_LR",
+   "scheduler_step_size" : 500,
+   "scheduler_gamma" : 0.75
+ },
+
+ "shortcut_prob_corr_factor" : {
+  "__comment" : "parameters for the shortcut for porb_corr_factor",
+  "values" : [0.4, 0.0],
+  "times" : [50, 150]
+ },
+
+ "soft_constraint" : {
+
+  "__comment": "all the parameters about the soft constraints",
+
+  "overlap": {
+    "__comment": "cost which discourages masks from overlapping",
+    "strength": 0,
+    "exponent": 1
+  },
+
+  "mask_volume_absolute" : {
+   "__comment" : "cost which discourage masks which are too large or too small",
+   "lower_bound_value": 40,
+   "lower_bound_width": 5,
+   "lower_bound_strength": 0,
+   "lower_bound_exponent": 2,
+   "upper_bound_value": 300,
+   "upper_bound_width": 5,
+   "upper_bound_strength": 0,
+   "upper_bound_exponent": 2
+  }
+ }
+}
+

dataset_multiMNIST/ML_parameters_WITH_GRID.json

@@ -0,0 +1,112 @@
+{
+
+ "neptune_project" : "dalessioluca/MNIST",
+
+ "simulation" : {
+   "__comment" : "there are 3 types of runs: scratch, resume, pretrained",
+   "type" : "scratch",
+   "MAX_EPOCHS" : 1501,
+   "TEST_FREQUENCY" : 50,
+   "CHECKPOINT_FREQUENCY" : 50,
+   "batch_size" : 128
+ },
+
+ "architecture" : {
+   "__comment" : "architecture parameters, level_zwhere_output is between 0 and n_max_pool included",
+   "dim_zinstance" : 20,
+   "dim_zwhere" : 4,
+   "dim_zbg" : 20,
+   "dim_logit" : 1,
+   "cropped_size" : 28,
+   "n_max_pool" : 4,
+   "level_zwhere_and_logit_output" : 2,
+   "level_background_output" : 4,
+   "n_ch_output_features" : 32,
+   "n_ch_after_first_two_conv" : 32,
+   "concatenate_raw_image_to_fmap" : true
+ },
+
+  "DPP" : {
+    "__comment" : "parameters for the DPP prior",
+    "n_kernels" : 1
+  },
+
+ "input_image" : {
+   "__comment" : "parameters describing the input images",
+   "n_objects_max" : 10,
+   "size_object_min" : 15,
+   "size_object_max" : 50,
+   "length_scale_similarity" : 10.0,
+   "size_raw_image" : 80,
+   "ch_in" : 1
+ },
+
+ "nms" : {
+   "__comment" : "parameters for the non-max-suppression",
+   "overlap_threshold_train" : 0.3,
+   "overlap_threshold_test" : 0.5
+ },
+
+ "GECO_loss" : {
+   "__comment" : "if active=false use ELBO, else use GECO with Log-Likelihood threshold = n_pixels * n_channel * threshold",
+   "is_active" : true,
+
+   "geco_mse_range" : [0.1, 0.8, 0.9],
+   "geco_fgfraction_range" : [0.1, 1.0, 10.0],
+   "geco_ncell_range" : [0.1, 1.0, 10.0],
+
+   "target_fgfraction" : [0.05, 0.15],
+   "target_mse" : [0.25, 0.75],
+   "target_ncell" : [3.0, 5.0],
+
+   "bg_std" : 0.1,
+   "fg_std" : 0.1
+ },
+
+ "optimizer" : {
+   "__comment" : "which optimizer to use",
+   "type" : "adam",
+   "base_lr" : 0.001,
+   "betas" : [0.9, 0.999],
+   "base_lr_similarity" : 0.01,
+   "betas_similarity" : [0.9, 0.999],
+   "base_lr_geco" : 0.001,
+   "betas_geco" : [0.9, 0.999],
+   "weight_decay" : 0.0,
+   "eps" : 1E-8,
+   "scheduler_is_active" : true,
+   "scheduler_type" : "step_LR",
+   "scheduler_step_size" : 500,
+   "scheduler_gamma" : 0.75
+ },
+
+ "shortcut_prob_corr_factor" : {
+  "__comment" : "parameters for the shortcut for porb_corr_factor",
+  "values" : [0.4, 0.0],
+  "times" : [50, 150]
+ },
+
+ "soft_constraint" : {
+
+  "__comment": "all the parameters about the soft constraints",
+
+  "overlap": {
+    "__comment": "cost which discourages masks from overlapping",
+    "strength": 0,
+    "exponent": 1
+  },
+
+  "mask_volume_absolute" : {
+   "__comment" : "cost which discourage masks which are too large or too small",
+   "lower_bound_value": 40,
+   "lower_bound_width": 5,
+   "lower_bound_strength": 0,
+   "lower_bound_exponent": 2,
+   "upper_bound_value": 300,
+   "upper_bound_width": 5,
+   "upper_bound_strength": 0,
+   "upper_bound_exponent": 2
+  }
+ }
+}
+