CSV file <- PyTorch state_dict key (TransformerNoConv, 3 encoders, best_model.pt epoch 106)
patch_weight.csv                    <- embed.weight  (128, 32)
patch_bias.csv                      <- embed.bias  (128,)
tokens.csv                          <- cls  (1, 1, 128)
position.csv                        <- pos  (1, 9, 128)
final_norm_weight.csv               <- norm.weight  (128,)
final_norm_bias.csv                 <- norm.bias  (128,)
classifier_weight.csv               <- fc.weight  (11, 128)
classifier_bias.csv                 <- fc.bias  (11,)
encoder_0/in_proj_weight.csv        <- encoder.layers.0.self_attn.in_proj_weight  (384, 128)
encoder_0/in_proj_bias.csv          <- encoder.layers.0.self_attn.in_proj_bias  (384,)
encoder_0/out_proj_weight.csv       <- encoder.layers.0.self_attn.out_proj.weight  (128, 128)
encoder_0/out_proj_bias.csv         <- encoder.layers.0.self_attn.out_proj.bias  (128,)
encoder_0/linear1_weight.csv        <- encoder.layers.0.linear1.weight  (512, 128)
encoder_0/linear1_bias.csv          <- encoder.layers.0.linear1.bias  (512,)
encoder_0/linear2_weight.csv        <- encoder.layers.0.linear2.weight  (128, 512)
encoder_0/linear2_bias.csv          <- encoder.layers.0.linear2.bias  (128,)
encoder_0/norm1_weight.csv          <- encoder.layers.0.norm1.weight  (128,)
encoder_0/norm1_bias.csv            <- encoder.layers.0.norm1.bias  (128,)
encoder_0/norm2_weight.csv          <- encoder.layers.0.norm2.weight  (128,)
encoder_0/norm2_bias.csv            <- encoder.layers.0.norm2.bias  (128,)
encoder_1/in_proj_weight.csv        <- encoder.layers.1.self_attn.in_proj_weight  (384, 128)
encoder_1/in_proj_bias.csv          <- encoder.layers.1.self_attn.in_proj_bias  (384,)
encoder_1/out_proj_weight.csv       <- encoder.layers.1.self_attn.out_proj.weight  (128, 128)
encoder_1/out_proj_bias.csv         <- encoder.layers.1.self_attn.out_proj.bias  (128,)
encoder_1/linear1_weight.csv        <- encoder.layers.1.linear1.weight  (512, 128)
encoder_1/linear1_bias.csv          <- encoder.layers.1.linear1.bias  (512,)
encoder_1/linear2_weight.csv        <- encoder.layers.1.linear2.weight  (128, 512)
encoder_1/linear2_bias.csv          <- encoder.layers.1.linear2.bias  (128,)
encoder_1/norm1_weight.csv          <- encoder.layers.1.norm1.weight  (128,)
encoder_1/norm1_bias.csv            <- encoder.layers.1.norm1.bias  (128,)
encoder_1/norm2_weight.csv          <- encoder.layers.1.norm2.weight  (128,)
encoder_1/norm2_bias.csv            <- encoder.layers.1.norm2.bias  (128,)
encoder_2/in_proj_weight.csv        <- encoder.layers.2.self_attn.in_proj_weight  (384, 128)
encoder_2/in_proj_bias.csv          <- encoder.layers.2.self_attn.in_proj_bias  (384,)
encoder_2/out_proj_weight.csv       <- encoder.layers.2.self_attn.out_proj.weight  (128, 128)
encoder_2/out_proj_bias.csv         <- encoder.layers.2.self_attn.out_proj.bias  (128,)
encoder_2/linear1_weight.csv        <- encoder.layers.2.linear1.weight  (512, 128)
encoder_2/linear1_bias.csv          <- encoder.layers.2.linear1.bias  (512,)
encoder_2/linear2_weight.csv        <- encoder.layers.2.linear2.weight  (128, 512)
encoder_2/linear2_bias.csv          <- encoder.layers.2.linear2.bias  (128,)
encoder_2/norm1_weight.csv          <- encoder.layers.2.norm1.weight  (128,)
encoder_2/norm1_bias.csv            <- encoder.layers.2.norm1.bias  (128,)
encoder_2/norm2_weight.csv          <- encoder.layers.2.norm2.weight  (128,)
encoder_2/norm2_bias.csv            <- encoder.layers.2.norm2.bias  (128,)

encoder_i/Wq.csv, Wk.csv, Wv.csv (128x128) = rows 0-127, 128-255, 256-383 of in_proj_weight
encoder_i/bq.csv, bk.csv, bv.csv (128)     = the same parts of in_proj_bias
