Prediction accuracy using earthformer is lower than the rainformer issue

[Helomin]

Both use the sevir dataset to predict for two hours
Using the adamw optimizer with a learning rate of 1e-4, none of them used the task learning rate optimization strategy
Here are the earthformer model parameter settings:
base_units=128,
block_units=None,
scale_alpha=1.0,
num_heads=4,
attn_drop=0.0,
proj_drop=0.0,
ffn_drop=0.0,

inter-attn downsample/upsample

downsample=2,
downsample_type='patch_merge',
upsample_type="upsample",
upsample_kernel_size=3,

encoder

enc_depth=[2, 2],
enc_attn_patterns=None,
enc_cuboid_size=[(4, 4, 4), (4, 4, 4)],
enc_cuboid_strategy=[('l', 'l', 'l'), ('d', 'd', 'd')],
enc_shift_size=[(0, 0, 0), (0, 0, 0)],
enc_use_inter_ffn=True,

decoder

dec_depth=[2, 2],
dec_cross_start=0,
dec_self_attn_patterns=None,
dec_self_cuboid_size=[(4, 4, 4), (4, 4, 4)],
dec_self_cuboid_strategy=[('l', 'l', 'l'), ('d', 'd', 'd')],
dec_self_shift_size=[(1, 1, 1), (0, 0, 0)],
dec_cross_attn_patterns=None,
dec_cross_cuboid_hw=[(4, 4), (4, 4)],
dec_cross_cuboid_strategy=[('l', 'l', 'l'), ('d', 'l', 'l')],
dec_cross_shift_hw=[(0, 0), (0, 0)],
dec_cross_n_temporal=[1, 2],
dec_cross_last_n_frames=None,
dec_use_inter_ffn=True,
dec_hierarchical_pos_embed=False,

global vectors

num_global_vectors=8,
use_dec_self_global=False,
dec_self_update_global=True,
use_dec_cross_global=False,
use_global_vector_ffn=False,
use_global_self_attn=True,
separate_global_qkv=True,
global_dim_ratio=1,
z_init_method='zeros',

# initial downsample and final upsample

initial_downsample_type="stack_conv",
initial_downsample_activation="leaky",

initial_downsample_type=="conv"

initial_downsample_scale=1,
initial_downsample_conv_layers=2,
final_upsample_conv_layers=2,

initial_downsample_type == "stack_conv"

initial_downsample_stack_conv_num_layers=3,
initial_downsample_stack_conv_dim_list=[16, 64, 128], # [96, 384, 768]
initial_downsample_stack_conv_downscale_list=[3, 2, 2],
initial_downsample_stack_conv_num_conv_list=[2, 2, 2],

# end of initial downsample and final upsample

ffn_activation='gelu',
gated_ffn=False,
norm_layer='layer_norm',
padding_type='ignore',
pos_embed_type='t+hw',
checkpoint_level=0,
use_relative_pos=True,
self_attn_use_final_proj=True,
dec_use_first_self_attn=False,

initialization

attn_linear_init_mode="0",
ffn_linear_init_mode="0",
conv_init_mode="0",
down_up_linear_init_mode="0",
norm_init_mode="0",