Skip to content

Commit

Permalink
checked in a bunch of yaml files
Browse files Browse the repository at this point in the history
  • Loading branch information
Ian Goodfellow committed Aug 12, 2011
1 parent 243a34c commit af68741
Show file tree
Hide file tree
Showing 14 changed files with 450 additions and 0 deletions.
36 changes: 36 additions & 0 deletions exploring_estimation_criteria/cifar_grbm_nce.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,36 @@

!obj:pylearn2.scripts.train.Train {
"dataset": &training_data !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
"model": !obj:pylearn2.models.normalized_ebm.NormalizedEBM {
"init_logZ" : 0.,
"learn_logZ" : 1,
"logZ_lr_scale" : .001,
"ebm" : !obj:pylearn2.models.rbm.GaussianBinaryRBM {
"nvis" : 192,
"nhid" : 400,
"irange" : 0.05,
"energy_function_class" : !obj:pylearn2.energy_functions.rbm_energy.grbm_type_1 {},
"learn_sigma" : True,
"init_sigma" : .4,
"init_bias_hid" : -2.,
"mean_vis" : False,
"sigma_lr_scale" : 1e-3
}
},
"algorithm": !obj:pylearn2.training_algorithms.sgd.SGD {
"learning_rate" : 1e-3,
"batch_size" : 50,
"batches_per_iter" : 200,
"monitoring_batches" : 2,
"monitoring_dataset" : !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
"cost" : !obj:pylearn2.costs.ebm_estimation.NCE {
"noise" : !obj:pylearn2.distributions.mnd.fit {
"dataset" : *training_data,
"n_samples" : 2000000
}
}
},
"save_path": "cifar_grbm_nce.pkl"
}


38 changes: 38 additions & 0 deletions exploring_estimation_criteria/cifar_grbm_nce_sphere.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,38 @@

!obj:pylearn2.scripts.train.Train {
"dataset": &training_data !pkl: "/data/lisatmp/goodfeli/cifar10_sphere_train_2M.pkl",
"model": !obj:pylearn2.models.normalized_ebm.NormalizedEBM {
"init_logZ" : 0.,
"learn_logZ" : 1,
"logZ_lr_scale" : .001,
"ebm" : !obj:pylearn2.models.rbm.GaussianBinaryRBM {
"nvis" : 192,
"nhid" : 400,
"irange" : 0.05,
"energy_function_class" : !obj:pylearn2.energy_functions.rbm_energy.grbm_type_1 {},
"learn_sigma" : True,
"init_sigma" : .2,
"init_bias_hid" : -2.,
"mean_vis" : False,
"sigma_lr_scale" : 1e-3
}
},
"algorithm": !obj:pylearn2.training_algorithms.sgd.SGD {
"learning_rate" : 1e-3,
"batch_size" : 50,
"batches_per_iter" : 200,
"monitoring_batches" : 2,
"monitoring_dataset" : !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
"cost" : !obj:pylearn2.costs.ebm_estimation.NCE {
"noise" : !obj:pylearn2.distributions.uniform_hypersphere.UniformHypersphere {
"radius": 1.,
"dim" : 192
},
"noise_per_clean" : 10
},
"learning_rate_adjuster" : !obj:pylearn2.training_algorithms.sgd.MonitorBasedLRAdjuster {}
},
"save_path": "cifar_grbm_nce_sphere.pkl"
}


31 changes: 31 additions & 0 deletions exploring_estimation_criteria/cifar_grbm_smd.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@

!obj:pylearn2.scripts.train.Train {
"dataset": !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
"model": !obj:pylearn2.models.rbm.GaussianBinaryRBM {
"nvis" : 192,
"nhid" : 400,
"irange" : 0.05,
"energy_function_class" : !obj:pylearn2.energy_functions.rbm_energy.grbm_type_1 {},
"learn_sigma" : True,
"init_sigma" : .4,
"init_bias_hid" : -2.,
"mean_vis" : False,
"sigma_lr_scale" : 1e-3

},
"algorithm": !obj:pylearn2.training_algorithms.sgd.SGD {
"learning_rate" : 1e-1,
"batch_size" : 5,
"batches_per_iter" : 2000,
"monitoring_batches" : 20,
"monitoring_dataset" : !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
"cost" : !obj:pylearn2.costs.ebm_estimation.SMD {
"corruptor": !obj:pylearn2.corruption.GaussianCorruptor {
"stdev": .4
}
}
},
"save_path": "cifar_grbm_smd.pkl"
}


34 changes: 34 additions & 0 deletions exploring_estimation_criteria/cifar_grbm_smd_one.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,34 @@

!obj:pylearn2.scripts.train.Train {
"dataset": &data !obj:pylearn2.datasets.dense_design_matrix.from_dataset {
"dataset" : !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
"num_examples" : 1
},
"model": !obj:pylearn2.models.rbm.GaussianBinaryRBM {
"nvis" : 192,
"nhid" : 400,
"irange" : 0.05,
"energy_function_class" : !obj:pylearn2.energy_functions.rbm_energy.grbm_type_1 {},
"learn_sigma" : False,
"init_sigma" : .1,
"mean_vis" : False,
"sigma_lr_scale" : 1e-3

},
"algorithm": !obj:pylearn2.training_algorithms.sgd.SGD {
"learning_rate" : 1e-3,
"batch_size" : 1,
"batches_per_iter" : 10000,
"monitoring_batches" : 1000,
"monitoring_dataset" : *data,
"cost" : !obj:pylearn2.costs.ebm_estimation.SMD {
"corruptor": !obj:pylearn2.corruption.GaussianCorruptor {
"stdev": .1
}
}
},
"save_path": "cifar_grbm_smd_one.pkl",
"callbacks": [!obj:fuck_you.FuckYouCallback {}]
}


33 changes: 33 additions & 0 deletions exploring_estimation_criteria/cifar_reconsE_smd.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,33 @@

!obj:pylearn2.scripts.train.Train {
"dataset": !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
"model": !obj:galatea.models.febm.FEBM {
"energy_function": !obj:galatea.energy_functions.scratch.recons_model_1 {
"nvis" : 192,
"nhid" : 400,
"irange" : .05,
"init_bias_hid" : 0.,

"init_vis_prec" : 1.,
"vis_prec_lr_scale" : .001,
"learn_vis_prec" : 1.,

"init_delta" : 0.0
}
},
"algorithm": !obj:pylearn2.training_algorithms.sgd.SGD {
"learning_rate" : 1e-3,
"batch_size" : 50,
"batches_per_iter" : 100,
"monitoring_batches" : 10,
"monitoring_dataset" : !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
"cost" : !obj:pylearn2.costs.ebm_estimation.SMD {
"corruptor": !obj:pylearn2.corruption.GaussianCorruptor {
"stdev": .3
}
}
},
"save_path": "cifar_reconsE_smd.pkl"
}


29 changes: 29 additions & 0 deletions exploring_estimation_criteria/cos_reconsE_sm_tiny.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,29 @@

!obj:pylearn2.scripts.train.Train {
"dataset": !pkl: "tiny_dataset.pkl",
"model": !obj:galatea.models.febm.FEBM {
"energy_function": !obj:galatea.energy_functions.scratch.recons_model_1 {
"nvis" : 2,
"nhid" : 10,
"irange" : 1.5,
"init_bias_hid" : 0.0,

"init_vis_prec" : 5.,
"vis_prec_lr_scale" : .001,
"learn_vis_prec" : 1.,

"init_delta" : 0.0
}
},
"algorithm": !obj:pylearn2.training_algorithms.sgd.SGD {
"learning_rate" : .002,
"batch_size" : 5,
"batches_per_iter" : 1000,
"monitoring_batches" : 10,
"monitoring_dataset" : !obj:pylearn2.datasets.cos_dataset.CosDataset {},
"cost" : !obj:pylearn2.costs.ebm_estimation.SM {}
},
"save_path": "cos_reconsE_sm_tiny.pkl"
}


33 changes: 33 additions & 0 deletions exploring_estimation_criteria/cos_reconsE_smd_tiny.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,33 @@

!obj:pylearn2.scripts.train.Train {
"dataset": !pkl: "tiny_dataset.pkl",
"model": !obj:galatea.models.febm.FEBM {
"energy_function": !obj:galatea.energy_functions.scratch.recons_model_1 {
"nvis" : 2,
"nhid" : 400,
"irange" : 1.8,
"init_bias_hid" : 0.,

"init_vis_prec" : 1.,
"vis_prec_lr_scale" : .001,
"learn_vis_prec" : 1.,

"init_delta" : 0.0
}
},
"algorithm": !obj:pylearn2.training_algorithms.sgd.SGD {
"learning_rate" : 1e-7,
"batch_size" : 5,
"batches_per_iter" : 100000,
"monitoring_batches" : 1000,
"monitoring_dataset" : !pkl: "tiny_dataset.pkl",
"cost" : !obj:pylearn2.costs.ebm_estimation.SMD {
"corruptor": !obj:pylearn2.corruption.GaussianCorruptor {
"stdev": 1.
}
}
},
"save_path": "cos_reconsE_smd_tiny.pkl"
}


38 changes: 38 additions & 0 deletions exploring_estimation_criteria/experiment_b.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,38 @@
#adjustment to experiment a
#experiment a seemed to get the gradient direction fairly stable throughout learning
#however, the gradient magnitude and the pdf itself were unstable
#at some point in learning the right side of the curve had mass but the left did not
#later in learning this switched
#I'm trying to compensate for this by using a larger batch size

!obj:pylearn2.scripts.train.Train {
"dataset": !obj:pylearn2.datasets.cos_dataset.CosDataset {},
"model": !obj:galatea.models.local_noise_ebm.LocalNoiseEBM {
"nvis" : 2,
"nhid" : 5,
"init_bias_hid" : 0.0,
"irange" : 5.0,
"init_noise_var" : 1.0,
"min_misclass" : .05,
"max_misclass" : .5,
"noise_var_scale_up" : 1.001,
"noise_var_scale_down" : .999,
"max_noise_var" : 2.,
"time_constant" : .1,
"learning_rate" : .002,
"different_examples" : 0.,
"init_vis_prec" : 30.,
"learn_vis_prec" : 1.,
"energy_function" : "mse autoencoder",
"init_delta" : -0.5
},
"algorithm": !obj:pylearn2.training_algorithms.default.DefaultTrainingAlgorithm {
"batch_size" : 100,
"batches_per_iter" : 50,
"monitoring_batches" : 10,
"monitoring_dataset" : !obj:pylearn2.datasets.cos_dataset.CosDataset {},

},
"save_path": "experiment_b.pkl"
}

33 changes: 33 additions & 0 deletions exploring_estimation_criteria/experiment_c.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,33 @@
#this learns a model with the right gradient direction everywhere, but totally wrong global structure

!obj:pylearn2.scripts.train.Train {
"dataset": !pkl: "finite_sample_dataset.pkl" ,
"model": !obj:galatea.models.local_noise_ebm.LocalNoiseEBM {
"nvis" : 2,
"nhid" : 400,
"init_bias_hid" : 0.0,
"irange" : 5.0,
"init_noise_var" : 1.0,
"min_misclass" : .05,
"max_misclass" : .5,
"noise_var_scale_up" : 1.001,
"noise_var_scale_down" : .999,
"max_noise_var" : 2.,
"time_constant" : .1,
"learning_rate" : .002,
"different_examples" : 0.,
"init_vis_prec" : 30.,
"learn_vis_prec" : 1.,
"energy_function" : "mse autoencoder",
"init_delta" : -0.5
},
"algorithm": !obj:pylearn2.training_algorithms.default.DefaultTrainingAlgorithm {
"batch_size" : 5,
"batches_per_iter" : 1000,
"monitoring_batches" : 10,
"monitoring_dataset" : !obj:pylearn2.datasets.cos_dataset.CosDataset {},

},
"save_path": "experiment_c.pkl"
}

31 changes: 31 additions & 0 deletions s3c/attempt_004.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@
#first attempt at using gradient update in the M step
!obj:pylearn2.scripts.train.Train {
"dataset": !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
"model": !obj:galatea.s3c.s3c.S3C {
"nvis" : 192,
"nhid" : 300,
"init_bias_hid" : -1.5,
"irange" : .5,
"init_B" : 3.,
"min_B" : 3.,
"max_B" : 10.,
"init_alpha" : 1.,
"min_alpha" : 1.,
"max_alpha" : 1000.,
"init_mu" : 5.,
"N_schedule" : [1.,2.,4.,8.,16.,32.,64.,128.,256.,300.],
"new_stat_coeff" : .01,
"learn_after" : 10000,
"m_step" : !obj:galatea.s3c.s3c.VHS_Grad_M_Step {
"learning_rate" : .001
}
},
"algorithm": !obj:pylearn2.training_algorithms.default.DefaultTrainingAlgorithm {
"batch_size" : 50,
"batches_per_iter" : 10,
"monitoring_batches" : 1,
"monitoring_dataset" : !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
},
"save_path": "attempt_004.pkl"
}

31 changes: 31 additions & 0 deletions s3c/attempt_005.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@
#first attempt at using M step that takes Q(U) into account
!obj:pylearn2.scripts.train.Train {
"dataset": !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
"model": !obj:galatea.s3c.s3c.S3C {
"nvis" : 192,
"nhid" : 300,
"init_bias_hid" : -1.5,
"irange" : .5,
"init_B" : 3.,
"min_B" : 3.,
"max_B" : 10.,
"init_alpha" : 1.,
"min_alpha" : 1.,
"max_alpha" : 1000.,
"init_mu" : 5.,
"N_schedule" : [1.,2.,4.,8.,16.,32.,64.,128.,256.,300.],
"new_stat_coeff" : .01,
"learn_after" : 10000,
"m_step" : !obj:galatea.s3c.s3c.VHSU_Grad_M_Step {
"learning_rate" : 1e-5
}
},
"algorithm": !obj:pylearn2.training_algorithms.default.DefaultTrainingAlgorithm {
"batch_size" : 50,
"batches_per_iter" : 10,
"monitoring_batches" : 1,
"monitoring_dataset" : !pkl: "/data/lisatmp/goodfeli/cifar10_preprocessed_train_2M.pkl",
},
"save_path": "attempt_005.pkl"
}

Loading

0 comments on commit af68741

Please sign in to comment.