Good work! However, I am confused about the parameter settings on the T31TFM_1618 dataset. I tried these parameters:
architecture: TSViT
batch_size: 16
dim: 256
dim_head: 64
dropout: 0.0
emb_dropout: 0.0
heads: 4
ignore_background: true
img_res: 48
max_seq_len: 16
num_channels: 14
num_classes: 20
num_features: 16
patch_size: 2
pool: cls
scale_dim: 4
spatial_depth: 4
temporal_depth: 4
class_weights: null
loss_function: masked_cross_entropy
lr_base: 1e-3
lr_min: 5e-6
lr_scheduler: cosine
lr_start: 1e-8
num_cycles: 1
num_epochs: 100
num_warmup_epochs: 10
reset_lr: true
steps: (0, 80000)
weight_decay: 0.0
But the IoU is only 62%, not 63.1% presented in the paper[ViTs for SITS]. Have I missed any parameters? Do the parameters mentioned above need to be adjusted?
Good work! However, I am confused about the parameter settings on the T31TFM_1618 dataset. I tried these parameters:
architecture: TSViT
batch_size: 16
dim: 256
dim_head: 64
dropout: 0.0
emb_dropout: 0.0
heads: 4
ignore_background: true
img_res: 48
max_seq_len: 16
num_channels: 14
num_classes: 20
num_features: 16
patch_size: 2
pool: cls
scale_dim: 4
spatial_depth: 4
temporal_depth: 4
class_weights: null
loss_function: masked_cross_entropy
lr_base: 1e-3
lr_min: 5e-6
lr_scheduler: cosine
lr_start: 1e-8
num_cycles: 1
num_epochs: 100
num_warmup_epochs: 10
reset_lr: true
steps: (0, 80000)
weight_decay: 0.0
But the IoU is only 62%, not 63.1% presented in the paper[ViTs for SITS]. Have I missed any parameters? Do the parameters mentioned above need to be adjusted?