صفحهٔ اصلی گشت‌و‌گذار راهنما
ورود
kekezack
/
X_SSL_Net
0
0
انشعاب 0
پرونده‌ها مشکلات 0 درخواست واکشی 0 ویکی
درخت: fbbbfe53c8
شاخه‌ها تگ‌ها
X_SSL_Net
/
tmp
/
VMamba-main
/
analyze
/
erf.py

erf.py 18 KB
تاريخچه خام

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184 
  1. # this is only a script !
    2. 

  2. if __name__ == "__main__":
    3. 

  3.     import os
    4. 

  4.     from utils import visualize, EffectiveReceiptiveField, BuildModels
    5. 

  5.     simpnorm = EffectiveReceiptiveField.simpnorm
    6. 

  6. 

  7.     HOME = os.environ["HOME"].rstrip("/")
    8. 

  8. 

  9.     showpath = os.path.join(os.path.dirname(os.path.abspath(__file__)), "./show").rstrip("/")
    10. 

  10.     data_path = "/media/Disk1/Dataset/ImageNet_ILSVRC2012"
    11. 

  11.     
    12. 

  12.     # main results
    13. 

  13.     if False:
    14. 

  14.         results_before = [
    15. 

  15.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_resnet_mmpretrain(only_backbone=True, scale="r50"), size=1024, data_path=data_path, norms=simpnorm), ""),
    16. 

  16.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_convnext(only_backbone=True, scale="tiny"), size=1024, data_path=data_path, norms=simpnorm), ""),
    17. 

  17.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_swin_mmpretrain(only_backbone=True, scale="tiny"), size=1024, data_path=data_path, norms=simpnorm), ""),
    18. 

  18.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_deit_mmpretrain(only_backbone=True, scale="small"), size=1024, data_path=data_path, norms=simpnorm), ""),
    19. 

  19.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_hivit_mmpretrain(only_backbone=True, scale="tiny", size=1024), size=1024, data_path=data_path, norms=simpnorm), ""),
    20. 

  20.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="tv2"), size=1024, data_path=data_path, norms=simpnorm), ""),
    21. 

  21.         ]
    22. 

  22.         results_after = [
    23. 

  23.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_resnet_mmpretrain(with_ckpt=True, only_backbone=True, scale="r50"), size=1024, data_path=data_path, norms=simpnorm), ""),
    24. 

  24.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_convnext(with_ckpt=True, only_backbone=True, scale="tiny"), size=1024, data_path=data_path, norms=simpnorm), ""),
    25. 

  25.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_swin_mmpretrain(with_ckpt=True, only_backbone=True, scale="tiny"), size=1024, data_path=data_path, norms=simpnorm), ""),
    26. 

  26.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_deit_mmpretrain(with_ckpt=True, only_backbone=True, scale="small"), size=1024, data_path=data_path, norms=simpnorm), ""),
    27. 

  27.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_hivit_mmpretrain(with_ckpt=True, only_backbone=True, scale="tiny", size=1024), size=1024, data_path=data_path, norms=simpnorm), ""),
    28. 

  28.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="tv2"), size=1024, data_path=data_path, norms=simpnorm), ""),
    29. 

  29.         ]
    30. 

  30.         visualize.visualize_snsmaps(results_before + results_after, savefig=f"{showpath}/erf_main.jpg", rows=2, sticks=False, figsize=(10, 10.75), cmap='RdYlGn')
    31. 

  31.     
    32. 

  32.     # erf for tv2, initv1, initv2; 
    33. 

  33.     if False:
    34. 

  34.         results_before = [
    35. 

  35.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="tv2"), size=1024, data_path=data_path, norms=simpnorm), ""),
    36. 

  36.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    37. 

  37.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d", ssm_init="v1"),
    38. 

  38.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    39. 

  39.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    40. 

  40.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d", ssm_init="v2"),
    41. 

  41.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    42. 

  42.         ]
    43. 

  43.         results_after = [
    44. 

  44.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="tv2"), size=1024, data_path=data_path, norms=simpnorm), ""),
    45. 

  45.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="flex", 
    46. 

  46.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d", ssm_init="v1"),
    47. 

  47.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/private/classification/vssms/vssm1_tiny_0230s_init1/ckpt_epoch_261.pth",
    48. 

  48.                 key="model_ema",
    49. 

  49.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    50. 

  50.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="flex", 
    51. 

  51.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d", ssm_init="v2"),
    52. 

  52.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/private/classification/vssms/vssm1_tiny_0230s_init2/ckpt_epoch_264.pth",
    53. 

  53.                 key="model_ema",
    54. 

  54.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    55. 

  55.         ]
    56. 

  56.         visualize.visualize_snsmaps(results_before + results_after, savefig=f"{showpath}/erf_initv0v1v2.jpg", rows=2, sticks=False, figsize=(10, 10.75), cmap='RdYlGn')
    57. 

  57. 

  58.     # erf for tv2, unidi, bidi, cas2d; ndw, unidindw, bidindw, cas2dndw; 
    59. 

  59.     if False:
    60. 

  60.         results_before = [
    61. 

  61.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="tv2"), size=1024, data_path=data_path, norms=simpnorm), ""),
    62. 

  62.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    63. 

  63.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v051d_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    64. 

  64.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    65. 

  65.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    66. 

  66.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v052d_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    67. 

  67.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    68. 

  68.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    69. 

  69.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v052dc_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    70. 

  70.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    71. 

  71.             
    72. 

  72.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    73. 

  73.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=-1, ssm_conv_bias=True, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    74. 

  74.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    75. 

  75.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    76. 

  76.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=-1, ssm_conv_bias=False, forward_type="v051d_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    77. 

  77.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    78. 

  78.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    79. 

  79.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=-1, ssm_conv_bias=False, forward_type="v052d_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    80. 

  80.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    81. 

  81.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    82. 

  82.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=-1, ssm_conv_bias=False, forward_type="v052dc_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    83. 

  83.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    84. 

  84.         ]
    85. 

  85.         results_after = [
    86. 

  86.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="tv2"), size=1024, data_path=data_path, norms=simpnorm), ""),
    87. 

  87.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    88. 

  88.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v051d_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    89. 

  89.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/ckpts/private/classification/vssms/vssm1_tiny_0230ab1d/ckpt_epoch_272.pth",
    90. 

  90.                 key="model_ema",
    91. 

  91.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    92. 

  92.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    93. 

  93.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v052d_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    94. 

  94.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/ckpts/private/classification/vssms/vssm1_tiny_0230ab2d/ckpt_epoch_269.pth",
    95. 

  95.                 key="model_ema",
    96. 

  96.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    97. 

  97.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    98. 

  98.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v052dc_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    99. 

  99.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/ckpts/private/classification/vssms/vssm1_tiny_0230ab2dc/ckpt_epoch_250.pth",
    100. 

  100.                 key="model_ema",
    101. 

  101.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    102. 

  102.             
    103. 

  103.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    104. 

  104.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=-1, ssm_conv_bias=True, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    105. 

  105.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/ckpts/private/classification/vssms/vssm1_tiny_0230s_ndw/ckpt_epoch_237.pth",
    106. 

  106.                 key="model_ema",
    107. 

  107.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    108. 

  108.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    109. 

  109.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=-1, ssm_conv_bias=False, forward_type="v051d_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    110. 

  110.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/ckpts/private/classification/vssms/vssm1_tiny_0230ab1d_ndw/ckpt_epoch_272.pth",
    111. 

  111.                 key="model_ema",
    112. 

  112.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    113. 

  113.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    114. 

  114.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=-1, ssm_conv_bias=False, forward_type="v052d_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    115. 

  115.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/ckpts/private/classification/vssms/vssm1_tiny_0230ab2d_ndw/ckpt_epoch_268.pth",
    116. 

  116.                 key="model_ema",
    117. 

  117.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    118. 

  118.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="flex", 
    119. 

  119.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=-1, ssm_conv_bias=False, forward_type="v052dc_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    120. 

  120.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/ckpts/private/classification/vssms/vssm1_tiny_0230ab2dc_ndw/ckpt_epoch_263.pth",
    121. 

  121.                 key="model_ema",
    122. 

  122.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    123. 

  123.         ]
    124. 

  124.         visualize.visualize_snsmaps(results_before + results_after, savefig=f"{showpath}/erf_scanmethod.jpg", rows=2, sticks=False, figsize=(10, 10.75), cmap='RdYlGn')
    125. 

  125. 

  126.     # erf for training stage for tv2
    127. 

  127.     if False:
    128. 

  128.         results_before = [
    129. 

  129.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="tv2"), size=1024, data_path=data_path, norms=simpnorm), ""),
    130. 

  130.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="flex", 
    131. 

  131.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    132. 

  132.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/private/vssm1_tiny_0230s/20240426174619/ckpt_epoch_2.pth",
    133. 

  133.                 key="model_ema",
    134. 

  134.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    135. 

  135.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="flex", 
    136. 

  136.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    137. 

  137.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/private/vssm1_tiny_0230s/20240426174619/ckpt_epoch_5.pth",
    138. 

  138.                 key="model_ema",
    139. 

  139.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    140. 

  140.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="flex", 
    141. 

  141.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    142. 

  142.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/private/vssm1_tiny_0230s/20240426174619/ckpt_epoch_8.pth",
    143. 

  143.                 key="model_ema",
    144. 

  144.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    145. 

  145.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="flex", 
    146. 

  146.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    147. 

  147.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/private/vssm1_tiny_0230s/20240426174619/ckpt_epoch_11.pth",
    148. 

  148.                 key="model_ema",
    149. 

  149.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    150. 

  150.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="flex", 
    151. 

  151.                 cfg=dict(dims=96, depths=[2,2,8,2], ssm_d_state=1, ssm_dt_rank="auto", ssm_ratio=1.0, ssm_conv=3, ssm_conv_bias=False, forward_type="v05_noz", mlp_ratio=4.0, downsample_version="v3", patchembed_version="v2", norm_layer="ln2d"),
    152. 

  152.                 ckpt=f"{HOME}/Workspace/PylanceAware/ckpts/private/vssm1_tiny_0230s/20240426174619/ckpt_epoch_29.pth",
    153. 

  153.                 key="model_ema",
    154. 

  154.             ), size=1024, data_path=data_path, norms=simpnorm), ""),
    155. 

  155.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="tv2"), size=1024, data_path=data_path, norms=simpnorm), ""),
    156. 

  156.         ]
    157. 

  157.         visualize.visualize_snsmaps(results_before, savefig=f"{showpath}/erf_trainprocess.jpg", rows=1, sticks=False, figsize=(10, 10.75), cmap='RdYlGn')
    158. 

  158. 

  159.     # erf for vim
    160. 

  160.     if True:
    161. 

  161.         from analyze_for_vim import ExtraDev
    162. 

  162.         results_before = [
    163. 

  163.             (EffectiveReceiptiveField.get_input_grad_avg(ExtraDev.build_vim_for_throughput(only_backbone=True, size=1024), size=1024, data_path=data_path, norms=simpnorm), ""),
    164. 

  164.         ]
    165. 

  165.         results_after = [
    166. 

  166.             (EffectiveReceiptiveField.get_input_grad_avg(ExtraDev.build_vim_for_throughput(with_ckpt=True, only_backbone=True, size=1024), size=1024, data_path=data_path, norms=simpnorm), ""),
    167. 

  167.         ]
    168. 

  168.         visualize.visualize_snsmaps(results_before + results_after, savefig=f"{showpath}/erf_vimmethods.jpg", rows=2, sticks=False, figsize=(10, 10.75), cmap='RdYlGn')
    169. 

  169. 

  170.     # erf for s4nd
    171. 

  171.     if True:
    172. 

  172.         results_before = [
    173. 

  173.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(only_backbone=True, scale="tv2"), size=1024, data_path=data_path, norms=simpnorm), ""),
    174. 

  174.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_s4nd(only_backbone=True, scale="ctiny"), size=1024, data_path=data_path, norms=simpnorm), ""),
    175. 

  175.         ]
    176. 

  176.         results_after = [
    177. 

  177.             (EffectiveReceiptiveField.get_input_grad_avg(BuildModels.build_vmamba(with_ckpt=True, only_backbone=True, scale="tv2"), size=1024, data_path=data_path, norms=simpnorm), ""),
    178. 

  178.             # (after_s4nd, ""),
    179. 

  179.         ]
    180. 

  180.         visualize.visualize_snsmaps(results_before + results_after, savefig=f"{showpath}/erf_s4ndmethods.jpg", rows=2, sticks=False, figsize=(10, 10.75), cmap='RdYlGn')
    181. 

  181. 

  182. 

  183. 

  184.