sonebu
moving over from github
5096607
Top:
epoch
state_dict
arch
optimizer_state_dict
optimizer_type
compression_sched
extras
-------------------------------------
arch: ai85simplenet
-------------------------------------
extras: {'current_top1': 55.76, 'best_top1': 55.76, 'best_epoch': 291, 'clipping_method': 'MAX_BIT_SHIFT'}
-------------------------------------
state_dict:
conv1
output_shift: [-0.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [8.]
bias_bits: [8.]
bias
total # of elements, shape: 16 , [16]
# of unique elements: 16
min, max, mean: -3141.0 , 6035.0 , 1187.375
weight
total # of elements, shape: 432 , [16, 3, 3, 3]
# of unique elements: 90
min, max, mean: -103.0 , 95.0 , -0.14583333
conv2
output_shift: [-1.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [4.]
bias_bits: [8.]
bias
total # of elements, shape: 20 , [20]
# of unique elements: 20
min, max, mean: -882.0 , 968.0 , 307.5
weight
total # of elements, shape: 2880 , [20, 16, 3, 3]
# of unique elements: 15
min, max, mean: -8.0 , 7.0 , -0.11736111
conv3
output_shift: [-2.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [2.]
bias_bits: [8.]
bias
total # of elements, shape: 20 , [20]
# of unique elements: 15
min, max, mean: -165.0 , 255.0 , 114.9
weight
total # of elements, shape: 3600 , [20, 20, 3, 3]
# of unique elements: 4
min, max, mean: -2.0 , 1.0 , -0.05888889
conv4
output_shift: [-2.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [2.]
bias_bits: [8.]
bias
total # of elements, shape: 20 , [20]
# of unique elements: 19
min, max, mean: -256.0 , 255.0 , 74.95
weight
total # of elements, shape: 3600 , [20, 20, 3, 3]
# of unique elements: 4
min, max, mean: -2.0 , 1.0 , -0.051944446
conv5
output_shift: [-2.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [2.]
bias_bits: [8.]
bias
total # of elements, shape: 20 , [20]
# of unique elements: 15
min, max, mean: -256.0 , 255.0 , 86.65
weight
total # of elements, shape: 3600 , [20, 20, 3, 3]
# of unique elements: 4
min, max, mean: -2.0 , 1.0 , -0.008611111
conv6
output_shift: [-2.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [2.]
bias_bits: [8.]
bias
total # of elements, shape: 20 , [20]
# of unique elements: 14
min, max, mean: -204.0 , 255.0 , 120.2
weight
total # of elements, shape: 3600 , [20, 20, 3, 3]
# of unique elements: 4
min, max, mean: -2.0 , 1.0 , -0.018055556
conv7
output_shift: [-2.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [2.]
bias_bits: [8.]
bias
total # of elements, shape: 44 , [44]
# of unique elements: 38
min, max, mean: -256.0 , 255.0 , 46.636364
weight
total # of elements, shape: 7920 , [44, 20, 3, 3]
# of unique elements: 4
min, max, mean: -2.0 , 1.0 , -0.036237374
conv8
output_shift: [-3.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [2.]
bias_bits: [8.]
bias
total # of elements, shape: 48 , [48]
# of unique elements: 16
min, max, mean: -256.0 , 255.0 , 83.666664
weight
total # of elements, shape: 19008 , [48, 44, 3, 3]
# of unique elements: 4
min, max, mean: -2.0 , 1.0 , -0.026936026
conv9
output_shift: [-3.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [2.]
bias_bits: [8.]
bias
total # of elements, shape: 48 , [48]
# of unique elements: 17
min, max, mean: -256.0 , 255.0 , 164.5625
weight
total # of elements, shape: 20736 , [48, 48, 3, 3]
# of unique elements: 4
min, max, mean: -2.0 , 1.0 , -0.042197146
conv10
output_shift: [-3.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [2.]
bias_bits: [8.]
bias
total # of elements, shape: 96 , [96]
# of unique elements: 34
min, max, mean: -256.0 , 255.0 , -113.583336
weight
total # of elements, shape: 41472 , [96, 48, 3, 3]
# of unique elements: 4
min, max, mean: -2.0 , 1.0 , 0.025583526
conv11
output_shift: [-1.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [4.]
bias_bits: [8.]
bias
total # of elements, shape: 512 , [512]
# of unique elements: 435
min, max, mean: -1024.0 , 1023.0 , -48.671875
[WARNING]: # of unique elements in bias tensor is more than that allowed by bias_bits.
This might be OK, since Maxim deployment repository right shifts these.
weight
total # of elements, shape: 49152 , [512, 96, 1, 1]
# of unique elements: 16
min, max, mean: -8.0 , 7.0 , -0.051452637
conv12
output_shift: [-2.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [4.]
bias_bits: [8.]
bias
total # of elements, shape: 128 , [128]
# of unique elements: 115
min, max, mean: -1024.0 , 1023.0 , 360.78125
weight
total # of elements, shape: 65536 , [128, 512, 1, 1]
# of unique elements: 16
min, max, mean: -8.0 , 7.0 , -0.100616455
conv13
output_shift: [-0.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [4.]
bias_bits: [8.]
bias
total # of elements, shape: 128 , [128]
# of unique elements: 111
min, max, mean: -1024.0 , 1023.0 , 317.4922
weight
total # of elements, shape: 147456 , [128, 128, 3, 3]
# of unique elements: 16
min, max, mean: -8.0 , 7.0 , 0.000386556
conv14
output_shift: [1.]
adjust_output_shift: [0.]
quantize_activation: [1.]
shift_quantile: [0.985]
weight bits: [4.]
bias_bits: [8.]
bias
total # of elements, shape: 100 , [100]
# of unique elements: 81
min, max, mean: -146.0 , 131.0 , -3.66
weight
total # of elements, shape: 12800 , [100, 128, 1, 1]
# of unique elements: 16
min, max, mean: -8.0 , 7.0 , -0.8530469