package fewshot.configs
Get desktop application:
View/edit binary Protocol Buffers messages
message
cnn_config.proto:7Used in:
optional int32 height = 1
optional int32 width = 2
optional int32 num_channel = 3
repeated int32 num_filters = 4
repeated int32 strides = 5
repeated string pool_fn = 6
repeated int32 pool_strides = 7
repeated string conv_act_fn = 8
message
experiment_config.proto:14optional string backbone_class = 1
optional string model_class = 2
optional string dtype = 3
optional resnet_config = 100
optional cnn_config = 101
optional protonet_config = 1000
optional transfer_config = 1001
optional optimizer_config = 2000
optional train_config = 3000
message
optimizer_config.proto:7Used in: ,
optional string optimizer = 1
Optimizer name.
repeated float lr_list = 2
Learning rate decay values.
repeated int32 lr_decay_steps = 3
Learning rate decay steps.
optional int32 max_train_steps = 4
Total number of training steps.
optional int32 batch_size = 5
Training batch size.
optional int32 num_gpu = 6
Training number of GPUs.
message
protonet_config.proto:7Used in:
optional string similarity = 1
Similarity function, euclidean or cosine.
optional bool cosine_a = 2
Whether to use cosine softmax in task A.
optional bool cosine_tau = 3
Whether to add a learnable scaling factor in task A.
optional bool cosine_softmax = 4
Whether to use cosine softmax in task B.
optional bool cosine_softmax_tau = 5
Whether to add a learnable scaling factor in task B
optional bool cosine_attention = 6
Whether to use cosine attention mechanism.
optional bool protos_phi = 7
Whether to gate the prototypes with learned parameters.
optional bool reinit_tau = 8
Whether to reinit tau variable for cosine A.
message
resnet_config.proto:7Used in:
optional int32 height = 1
Input image height.
optional int32 width = 2
Input image width.
optional int32 num_channel = 3
Input image number of channels.
repeated int32 num_residual_units = 4
Number of residual units.
repeated int32 num_filters = 5
Input image number of filters.
repeated int32 strides = 6
Stride for the initial convolution of each resolution stage.
optional int32 init_stride = 7
Initial convolution strides.
optional bool init_max_pool = 8
Whether doing max pooling in the initial convolution.
optional int32 init_filter = 9
Number of filters in the initial convolution.
optional bool use_bottleneck = 10
Whether to use bottleneck layer in each residual unit.
optional float wd = 11
Weight decay.
optional string normalization = 12
Normalization scheme used in every layer.
optional int32 num_norm_groups = 13
Number of groups (used by GroupNorm).
optional bool global_avg_pool = 14
Whether perform global average pooling in the end.
optional string data_format = 15
Data format, NCHW or NHWC.
optional string version = 16
ResNet version, v1, v2, snailv1, or snailv2.
optional float leaky_relu = 17
Whether to use Leaky ReLU, if yes, by how much.
optional string filter_initialization = 18
Filter initialization scheme, normal or uniform.
optional bool add_last_relu = 19
Add the final ReLU to the feature map.
message
train_config.proto:7Used in:
optional int32 steps_per_val = 1
Number of steps for one validation run.
optional int32 steps_per_log = 2
Number of steps per output printing.
optional int32 steps_per_save = 3
Number of steps per mode checkpointing.
optional int32 pretrain_a_steps = 4
Number of steps to pretrain task A.
message
transfer_config.proto:9Used in:
optional string finetune_layers = 1
Layers to finetune.
optional bool reduce_ft_lr = 2
Whether to reduce learning rate for lower layers.
optional float dummy_lr = 3
Dummy learning rate to build one-step computation graph.
optional float rbp_lambda = 4
Damping rate.
optional int32 rbp_nstep = 5
Number of steps of RBP loops.
optional int32 rbp_batch_size = 6
RBP batch size.
optional float finetune_wd = 7
Finetune weight decay.
optional bool old_and_new = 8
Whether the query set contains both old and new instances.
optional string transfer_loss_type = 9
Type of the additional transfer learning loss.
optional float transfer_loss_ratio = 10
Ratio of the transfer loss.
optional bool meta_only = 11
Whether to learn meta-weights only.
optional bool train_wclass_a = 12
Whether to finetune readout weights for task A.
optional bool replace_grad_wclass_a = 13
Replace the gradients for weights for task A to 1-step.
optional float reduce_slow_layers = 14
Fractional learning rate for slow layers
optional bool proto_init = 15
Initialize weights with prototypes.
optional bool cache_transfer_loss_var = 16
Cache the transfer loss values.
optional string scipy_interface = 17
Scipy optimizer interface.
optional string meta_layers = 18
Layers for meta-learning altogether.
optional int32 learned_memory_size = 19
Learned memory size.
optional float wd = 20
Meta weights weight decay.
optional bool warm_start = 21
Whether to warm start with no learned regularizer.
optional float attn_attr_tau_init = 22
Initial value for the cosine similarity scale variable.
optional float mlp_init = 23
Initial variance of the MLP.
optional string fast_model = 24
Fast model type, LR or MLP.
optional int32 fast_mlp_hidden = 25
Fast model MLP hidden size.
optional float init_gamma = 26
Initial log gamma.
optional bool learn_gamma = 27
Whether to learn gamma.
optional int32 bptt_steps = 28
Number of BPTT steps.
optional bool transient_bptt = 29
Whether to add BPTT in the transient phase.
optional bool steady_rbp = 30
Whether to add RBP in the steady phase.
optional bool bptt_all = 31
Whether to do BPTT at all steps.
optional float bptt_ratio = 32
Loss multiplier for bptt.
optional bool disable_rbp_embedding = 103
Disable RBP to backprop into embeddings.
optional int32 mlp_hidden = 100
MLP hidden size for the transfer loss function.
optional string mlp_act = 101
MLP activation function.
optional int32 mlp_layers = 102
MLP number layers for the transfer loss function.
optional float cost_a_ratio = 1000
Cost multiplier for task A.
optional float cost_b_ratio = 1001
Cost multiplier for task B.
optional float cost_remember = 1002
Cost to remember task A (not used).
optional int32 offset_a = 2000
Number of steps when the cool down of task A starts.
optional float final_cost_a_ratio = 2001
Final task A ratio.
optional int32 warmup_b = 2002
Number of steps to warm up B from 0.0.
optional int32 cooldown_a = 2003
Number of steps to cool down A.
optional ft_optimizer_config = 3000
Finetune optimizer config.