Important: This notebook will only work with fastai-0.7.x. Do not try to run any fastai-1.x code from this path in the repository because it will load fastai-0.7.x
AdamW/SGDW benchmarking¶
This is to benchmark an implementation of https://arxiv.org/abs/1711.05101
PATH = "/home/as/datasets/fastai.cifar10/cifar10/"
%matplotlib inline
%reload_ext autoreload
%autoreload 2
# This file contains all the main external libs we'll use
from fastai.imports import *
from fastai.transforms import *
from fastai.conv_learner import *
from fastai.model import *
from fastai.dataset import *
from fastai.sgdr import *
from fastai.plots import *
import matplotlib.pyplot as plt
%matplotlib inline
Common stuff for all experiments¶
def Get_SGD_Momentum(momentum=0.9):
return lambda *args, **kwargs: optim.SGD(*args, momentum=momentum, **kwargs)
def Get_Adam():
return lambda *args, **kwargs: optim.Adam(*args, **kwargs)
import pickle
def save_list(fname, l):
with open(fname, "wb") as fp:
pickle.dump(l, fp)
def read_list(fname):
with open(fname, "rb") as fp:
return pickle.load(fp)
Section 1: Plot loss trends of various scenarios¶
This is a common function which does the training.¶
The only thing it asks for is the optimizer, and the initial LR for that optimizer. Hence we are comparing optimizers keeping all things same.
def experiment(optimizer, PATH, lr=1e-3, find_lr=False, use_wd_sched=False, wds=None, do_unfreeze=False,
norm_wds=False, wds_sched_mult=None):
sz = 32
bs = 120
arch=resnet152
cycle_len=2
cycle_mult=2
num_cycles = 4
lr = lr
if wds is None:
weight_decay = 0.025 # As used in the paper https://arxiv.org/abs/1711.05101
else:
weight_decay = wds
tfms = tfms_from_model(arch, sz, aug_tfms=transforms_side_on, max_zoom=1.1)
data = ImageClassifierData.from_paths(PATH, tfms=tfms, bs=bs)
learn = ConvLearner.pretrained(arch, data, precompute=True, xtra_fc=[1024, 512], opt_fn=optimizer)
if find_lr:
lrf=learn.lr_find()
learn.sched.plot()
return
learn.fit(lr, 1, wds=weight_decay, use_wd_sched=use_wd_sched, norm_wds=norm_wds, wds_sched_mult=wds_sched_mult)
print('Now with precompute as False')
if do_unfreeze:
learn.unfreeze()
learn.precompute=False
learn.fit(lr, num_cycles, wds=weight_decay, use_wd_sched=use_wd_sched, cycle_len=cycle_len,
cycle_mult=cycle_mult, norm_wds=norm_wds, wds_sched_mult=wds_sched_mult)
loss = learn.sched.losses
fig = plt.figure(figsize=(10, 5))
plt.plot(loss)
plt.show()
learn.sched.plot_lr()
return learn.sched.losses, learn
f = lambda x: np.array(x.layer_opt.lrs) / x.init_lrs
SGDR/SGDW¶
%%time
sgdm = Get_SGD_Momentum()
loss_sgdm = experiment(sgdm, PATH=PATH, find_lr=True)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
88%|████████▊ | 367/417 [00:06<00:00, 53.78it/s, loss=6.84] CPU times: user 12.9 s, sys: 6.71 s, total: 19.6 s Wall time: 13.8 s
%%time
# Without weight decay
sgdm = Get_SGD_Momentum()
loss_sgdm, _ = experiment(sgdm, PATH=PATH, lr=1e-2)
save_list('sgdm_loss.txt', loss_sgdm)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.10252 1.04199 0.65605] Now with precompute as False
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.1364 1.02446 0.65794] [ 1. 1.08206 0.97455 0.6751 ] [ 2. 1.15686 1.07465 0.63869] [ 3. 1.15261 1.0353 0.65099] [ 4. 1.10498 0.98328 0.67708] [ 5. 1.06011 0.95787 0.68948] [ 6. 1.19456 1.09135 0.63075] [ 7. 1.19494 1.08135 0.63065] [ 8. 1.19586 1.06903 0.64256] [ 9. 1.16876 1.06267 0.65228] [ 10. 1.13828 1.02387 0.65903] [ 11. 1.11264 0.99049 0.67401] [ 12. 1.07651 0.97051 0.68056] [ 13. 1.06247 0.96359 0.68204] [ 14. 1.22855 1.07522 0.63919] [ 15. 1.21781 1.11106 0.63383] [ 16. 1.20853 1.08698 0.63849] [ 17. 1.22872 1.08815 0.6373 ] [ 18. 1.23351 1.10872 0.62689] [ 19. 1.19806 1.07182 0.64554] [ 20. 1.21998 1.06709 0.64554] [ 21. 1.18835 1.05453 0.64841] [ 22. 1.17914 1.04329 0.64752] [ 23. 1.17402 1.05659 0.64583] [ 24. 1.15148 1.02461 0.66141] [ 25. 1.15086 1.00453 0.66806] [ 26. 1.12338 0.9989 0.67123] [ 27. 1.11844 0.9805 0.67897] [ 28. 1.09065 0.97452 0.67867] [ 29. 1.08323 0.97258 0.68135]
CPU times: user 23min 22s, sys: 6min 45s, total: 30min 8s Wall time: 20min 7s
%%time
# With weight decay
sgdmw = Get_SGD_Momentum()
loss_sgdmw, _ = experiment(sgdmw, PATH=PATH, lr=1e-2, use_wd_sched=True, norm_wds=True, wds_sched_mult=f)
save_list('sgdmw_loss.txt', loss_sgdmw)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.09579 0.98664 0.65526] Now with precompute as False
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.12314 0.95745 0.66409] [ 1. 1.07291 0.93583 0.66964] [ 2. 1.0849 0.93277 0.67599] [ 3. 1.03306 0.91303 0.68145] [ 4. 1.01905 0.88995 0.69058] [ 5. 0.98558 0.89072 0.68849] [ 6. 1.02736 0.91853 0.67559] [ 7. 1.0194 0.90845 0.68353] [ 8. 0.99881 0.89167 0.6869 ] [ 9. 0.99231 0.86899 0.69524] [ 10. 0.94204 0.84914 0.70357] [ 11. 0.91504 0.83937 0.70585] [ 12. 0.89167 0.83442 0.70883] [ 13. 0.89308 0.83453 0.70982] [ 14. 0.98872 0.88613 0.69048] [ 15. 0.99674 0.88223 0.69593] [ 16. 0.97349 0.88411 0.69276] [ 17. 0.9857 0.87319 0.69782] [ 18. 0.96493 0.8684 0.69712] [ 19. 0.94969 0.85978 0.69901] [ 20. 0.94785 0.85255 0.70218] [ 21. 0.92961 0.84641 0.70992] [ 22. 0.90159 0.83513 0.70675] [ 23. 0.89571 0.83458 0.70863] [ 24. 0.87042 0.82623 0.71111] [ 25. 0.85979 0.81458 0.71696] [ 26. 0.82356 0.81444 0.71647] [ 27. 0.80551 0.81218 0.71875] [ 28. 0.80451 0.80461 0.71845] [ 29. 0.81771 0.80785 0.71935]
CPU times: user 23min 27s, sys: 6min 52s, total: 30min 19s Wall time: 20min 29s
Adam/AdamW¶
adam = Get_Adam()
loss_adam = experiment(adam, PATH, find_lr=True)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
75%|███████▍ | 311/417 [00:07<00:02, 42.19it/s, loss=5.74]
Train
%%time
# Without weight decay
adam = Get_Adam()
loss_adam, _ = experiment(adam, PATH, 1e-3)
save_list('adam_loss.txt', loss_adam)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.16206 1.10624 0.61597] Now with precompute as False
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.25446 1.11698 0.62788] [ 1. 1.17912 1.03288 0.65288] [ 2. 1.27068 1.13769 0.62192] [ 3. 1.24414 1.10389 0.63661] [ 4. 1.18874 1.03594 0.65635] [ 5. 1.1632 1.01193 0.6623 ] [ 6. 1.2639 1.14672 0.61895] [ 7. 1.2698 1.13121 0.61746] [ 8. 1.26611 1.10718 0.63443] [ 9. 1.22351 1.09626 0.63621] [ 10. 1.20549 1.05746 0.65238] [ 11. 1.17759 1.02645 0.66478] [ 12. 1.14593 1.0053 0.67024] [ 13. 1.1308 0.99884 0.67431] [ 14. 1.27133 1.1385 0.62113] [ 15. 1.24347 1.11354 0.63105] [ 16. 1.24738 1.11663 0.62867] [ 17. 1.27063 1.12279 0.6251 ] [ 18. 1.24126 1.10634 0.62282] [ 19. 1.25247 1.0985 0.63462] [ 20. 1.24079 1.08414 0.64811] [ 21. 1.22069 1.07312 0.64415] [ 22. 1.22488 1.07028 0.6499 ] [ 23. 1.20484 1.06734 0.65099] [ 24. 1.16112 1.03352 0.65913] [ 25. 1.16851 1.02562 0.66131] [ 26. 1.12653 1.00558 0.66667] [ 27. 1.12084 0.99636 0.67192] [ 28. 1.11384 0.98569 0.67599] [ 29. 1.10518 0.98719 0.67559]
CPU times: user 23min 18s, sys: 6min 59s, total: 30min 17s Wall time: 20min 23s
%%time
# With weight decay
adamw = Get_Adam()
loss_adamw, _ = experiment(adamw, PATH, 1e-3, use_wd_sched=True, norm_wds=True, wds_sched_mult=f)
save_list('adamw_loss.txt', loss_adamw)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.05115 0.96719 0.6628 ] Now with precompute as False
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.08162 0.94352 0.67183] [ 1. 0.99748 0.89464 0.6881 ] [ 2. 1.0483 0.91754 0.67679] [ 3. 1.02091 0.89119 0.68889] [ 4. 0.95075 0.85186 0.70327] [ 5. 0.90177 0.84354 0.70506] [ 6. 1.00654 0.89183 0.68552] [ 7. 0.99001 0.88197 0.69018] [ 8. 0.96011 0.87726 0.69534] [ 9. 0.96041 0.85465 0.70427] [ 10. 0.90791 0.84278 0.70377] [ 11. 0.87729 0.82705 0.71131] [ 12. 0.86129 0.81556 0.71369] [ 13. 0.84167 0.81839 0.71319] [ 14. 0.96958 0.88032 0.69177] [ 15. 0.96295 0.86398 0.69831] [ 16. 0.97039 0.87492 0.69117] [ 17. 0.9624 0.87117 0.69564] [ 18. 0.95494 0.86043 0.70437] [ 19. 0.93728 0.85608 0.70089] [ 20. 0.90854 0.84553 0.70397] [ 21. 0.90762 0.84097 0.70883] [ 22. 0.8828 0.83024 0.7127 ] [ 23. 0.85682 0.82416 0.71329] [ 24. 0.84861 0.80681 0.71984] [ 25. 0.81995 0.80552 0.71875] [ 26. 0.79154 0.79567 0.72321] [ 27. 0.78195 0.79831 0.72401] [ 28. 0.77546 0.79399 0.72698] [ 29. 0.76665 0.79755 0.72421]
CPU times: user 23min 25s, sys: 7min, total: 30min 26s Wall time: 20min 39s
Differential Learning and Weight Decay¶
%%time
adamw_diff = Get_Adam()
lr = 1e-3
wd = 0.025
loss_adamw_diff, _ = experiment(adamw_diff, PATH, [lr/10, lr/5, lr], wds=[wd/10, wd/5, wd], use_wd_sched=True, norm_wds=True, wds_sched_mult=f)
save_list('loss_adamw_diff.txt', loss_adamw_diff)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.25343 1.04188 0.63175] Now with precompute as False
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.12512 0.96615 0.66171] [ 1. 1.05414 0.92497 0.6746 ] [ 2. 1.06102 0.92863 0.67649] [ 3. 1.0196 0.90395 0.68185] [ 4. 0.9773 0.87001 0.69395] [ 5. 0.92542 0.85536 0.70188] [ 6. 1.0287 0.91618 0.67996] [ 7. 0.99776 0.89601 0.68482] [ 8. 0.95506 0.87979 0.69028] [ 9. 0.95912 0.85556 0.70357] [ 10. 0.9116 0.84213 0.70575] [ 11. 0.88703 0.82763 0.70784] [ 12. 0.84927 0.81854 0.71081] [ 13. 0.85036 0.81479 0.7128 ] [ 14. 0.98082 0.87497 0.69147] [ 15. 0.95896 0.87572 0.69266] [ 16. 0.95894 0.86377 0.69841] [ 17. 0.95678 0.86425 0.69365] [ 18. 0.93004 0.86835 0.69881] [ 19. 0.93132 0.86679 0.69871] [ 20. 0.9289 0.8438 0.70625] [ 21. 0.90753 0.84369 0.70675] [ 22. 0.86527 0.83652 0.71191] [ 23. 0.88072 0.81857 0.71349] [ 24. 0.84527 0.81106 0.71786] [ 25. 0.81415 0.80362 0.71935] [ 26. 0.77724 0.79865 0.71984] [ 27. 0.79264 0.79593 0.72143] [ 28. 0.77519 0.79751 0.72173] [ 29. 0.77812 0.79563 0.72093]
CPU times: user 23min 26s, sys: 7min, total: 30min 27s Wall time: 20min 38s
%%time
sgdw_diff = Get_SGD_Momentum(0.9)
lr = 1e-2
wd = 0.025
loss_sgdw_diff, _ = experiment(sgdw_diff, PATH, [lr/10, lr/5, lr], wds=[wd/10, wd/5, wd], use_wd_sched=True, norm_wds=True, wds_sched_mult=f)
save_list('loss_sgdw_diff.txt', loss_sgdw_diff)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.38557 1.14619 0.59564] Now with precompute as False
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.2172 1.00715 0.64316] [ 1. 1.14248 0.97364 0.65169] [ 2. 1.11695 0.95135 0.65883] [ 3. 1.07557 0.93486 0.67143] [ 4. 1.02138 0.90637 0.68105] [ 5. 1.00519 0.90298 0.67976] [ 6. 1.04351 0.92113 0.67302] [ 7. 1.02742 0.9078 0.67788] [ 8. 1.01408 0.88842 0.68651] [ 9. 0.99154 0.87992 0.68998] [ 10. 0.95499 0.86702 0.69603] [ 11. 0.95086 0.85147 0.70109] [ 12. 0.92738 0.84287 0.70387] [ 13. 0.91638 0.84244 0.70496] [ 14. 1.01557 0.90427 0.68561] [ 15. 1.0169 0.88434 0.68641] [ 16. 0.98096 0.87993 0.69355] [ 17. 0.9812 0.87104 0.69941] [ 18. 0.97249 0.85646 0.70218] [ 19. 0.9625 0.8644 0.69911] [ 20. 0.95107 0.84487 0.70268] [ 21. 0.92236 0.84518 0.70437] [ 22. 0.90926 0.83892 0.70308] [ 23. 0.88227 0.82956 0.70923] [ 24. 0.8835 0.81935 0.7126 ] [ 25. 0.85533 0.80991 0.71548] [ 26. 0.83115 0.80799 0.71954] [ 27. 0.81374 0.8034 0.71796] [ 28. 0.8122 0.80351 0.71806] [ 29. 0.79961 0.80023 0.71905]
CPU times: user 23min 20s, sys: 6min 57s, total: 30min 18s Wall time: 20min 27s
fig=plt.figure(figsize=(15, 10))
plt.plot(loss_adam, c='red', label='Adam')
plt.plot(loss_sgdm, c='blue', label='SGDM')
plt.plot(loss_adamw, c='green', label='AdamW')
plt.plot(loss_sgdmw, c='black', label='SGDW')
plt.plot(loss_adamw_diff, c='orange', label='AdamW_differential')
plt.plot(loss_sgdw_diff, c='gray', label='SGDW_differential')
plt.legend()
plt.show()
Section 2: Check for regularization of overfitting¶
import time
def check_overfitting(optimizer, PATH, sz, bs, lr, wds, use_wd_sched=True):
arch=resnet50
cycle_len=12
cycle_mult=2
num_cycles = 1
tfms = tfms_from_model(arch, sz) # aug_tfms=transforms_side_on, max_zoom=1.1
data = ImageClassifierData.from_paths(PATH, tfms=tfms, bs=bs)
# Turning off Dropout, adding 3 extra FC layers to help in overfitting.
learn = ConvLearner.pretrained(arch, data, precompute=False, xtra_fc=[1024, 512], ps=[0, 0, 0], opt_fn=optimizer)
print("==== Let's overfit the model. Expectation: training loss should reduce but validation loss should stagnate.")
learn.unfreeze()
learn.fit(lr, num_cycles, cycle_len=cycle_len, cycle_mult=cycle_mult)
print("==== Let's introduce weight regularization. Expectation: training loss and validation loss should reduce. Accuracy should improve.")
learn.fit(lr, num_cycles, wds=wds, use_wd_sched=use_wd_sched, cycle_len=cycle_len, cycle_mult=cycle_mult)
time.sleep(5)
AdamW on Cats & Dogs
%%time
PATH = "/home/as/datasets/fastai.dogscats/"
check_overfitting(Get_Adam(), PATH=PATH, sz=224, bs=96, lr=1e-3, wds=0.25, use_wd_sched=True)
AdamW on Cifar10
%%time
PATH = "/home/as/datasets/fastai.cifar10/cifar10/"
check_overfitting(Get_Adam(), PATH=PATH, sz=32, bs=320, lr=1e-3, wds=0.45, use_wd_sched=True)
Adam on Cifar10
%%time
PATH = "/home/as/datasets/fastai.cifar10/cifar10/"
check_overfitting(Get_Adam(), PATH=PATH, sz=32, bs=320, lr=1e-3, wds=0.45, use_wd_sched=False)
Section 3: Tests after splitting the various parts of the regularizer¶
Separate weight regularization
%%time
# With weight decay
adamw = Get_Adam()
lr = [1e-5,1e-4,1e-3]
wd = [1e-7,1e-6,1e-5]
loss_adamw1, learn1 = experiment(adamw, PATH, lr, use_wd_sched=True, wds=wd, do_unfreeze=True)
save_list('adamw_loss1.txt', loss_adamw1)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.88647 1.47695 0.51379] Now with precompute as False
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.18962 0.96506 0.65635] [ 1. 1.11035 0.92796 0.67093] [ 2. 1.07023 0.92718 0.67153] [ 3. 1.02732 0.89155 0.68552] [ 4. 0.97336 0.87203 0.69375] [ 5. 0.96949 0.87093 0.69494] [ 6. 1.02327 0.89319 0.68591] [ 7. 0.97858 0.87634 0.69097] [ 8. 0.97534 0.85293 0.70149] [ 9. 0.93149 0.84771 0.70268] [ 10. 0.92056 0.82565 0.7124 ] [ 11. 0.88824 0.82935 0.70972] [ 12. 0.87601 0.81957 0.7129 ] [ 13. 0.86438 0.81952 0.7121 ] [ 14. 0.94136 0.86304 0.69673] [ 15. 0.93918 0.83937 0.70506] [ 16. 0.92601 0.84418 0.70675] [ 17. 0.89236 0.82786 0.7131 ] [ 18. 0.89477 0.83408 0.71012] [ 19. 0.89333 0.81779 0.71607] [ 20. 0.84657 0.81528 0.71528] [ 21. 0.84046 0.81346 0.71409] [ 22. 0.83054 0.811 0.71845] [ 23. 0.81166 0.80624 0.71925] [ 24. 0.79839 0.80018 0.71974] [ 25. 0.79172 0.79752 0.72371] [ 26. 0.7638 0.80074 0.72292] [ 27. 0.76813 0.79384 0.72371] [ 28. 0.77779 0.79411 0.72361] [ 29. 0.76272 0.79376 0.7246 ]
CPU times: user 23min 17s, sys: 6min 57s, total: 30min 14s Wall time: 20min 22s
With weight normalization
%%time
# With weight decay
adamw = Get_Adam()
lr = [1e-5,1e-4,1e-3]
wd = [1e-5,1e-4,1e-3]
loss_adamw2, learn2 = experiment(adamw, PATH, lr, use_wd_sched=True, wds=wd, norm_wds=True, do_unfreeze=True)
save_list('adamw_loss2.txt', loss_adamw2)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.88573 1.45961 0.51915] Now with precompute as False
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.19242 0.96442 0.65804] [ 1. 1.09521 0.93045 0.66726] [ 2. 1.09251 0.9283 0.67371] [ 3. 1.01866 0.88867 0.68571] [ 4. 0.98586 0.87771 0.69256] [ 5. 0.96249 0.86489 0.69534] [ 6. 1.00961 0.89263 0.68571] [ 7. 0.99576 0.89136 0.68522] [ 8. 0.96364 0.86692 0.69415] [ 9. 0.93769 0.85321 0.70139] [ 10. 0.90633 0.83638 0.70476] [ 11. 0.88083 0.82757 0.70843] [ 12. 0.85543 0.8223 0.71171] [ 13. 0.86451 0.81844 0.7125 ] [ 14. 0.94469 0.86198 0.69732] [ 15. 0.91937 0.83931 0.70585] [ 16. 0.92957 0.84116 0.70228] [ 17. 0.89314 0.82789 0.71032] [ 18. 0.88034 0.83583 0.70823] [ 19. 0.87155 0.81551 0.71458] [ 20. 0.85164 0.81305 0.71577] [ 21. 0.84515 0.80522 0.71796] [ 22. 0.82017 0.80378 0.71944] [ 23. 0.80766 0.80399 0.72143] [ 24. 0.78294 0.79544 0.72341] [ 25. 0.79158 0.79872 0.72193] [ 26. 0.77365 0.79022 0.72421] [ 27. 0.76788 0.79124 0.7252 ] [ 28. 0.76734 0.79056 0.72331] [ 29. 0.76179 0.78912 0.7249 ]
CPU times: user 23min 18s, sys: 6min 54s, total: 30min 13s Wall time: 20min 22s
With custom weight multiplier wds_sched_mult
%%time
# With weight decay
adamw = Get_Adam()
lr = [1e-5,1e-4,1e-3]
wd = [1e-5,1e-4,1e-3]
f = lambda x: np.array(x.layer_opt.lrs) / x.init_lrs
loss_adamw3, learn3 = experiment(adamw, PATH, lr, use_wd_sched=True, wds=wd, norm_wds=True, wds_sched_mult=f, do_unfreeze=True)
save_list('adamw_loss3.txt', loss_adamw3)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.91495 1.45581 0.51587] Now with precompute as False
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.18653 0.97802 0.65179] [ 1. 1.08818 0.92862 0.66667] [ 2. 1.08374 0.92995 0.67014] [ 3. 1.02608 0.89471 0.68403] [ 4. 0.99638 0.86567 0.69564] [ 5. 0.95797 0.86039 0.69444] [ 6. 1.01389 0.89751 0.68373] [ 7. 0.98849 0.87242 0.69206] [ 8. 0.95054 0.85777 0.69434] [ 9. 0.92966 0.85104 0.7003 ] [ 10. 0.91 0.8263 0.71091] [ 11. 0.87937 0.82215 0.71349] [ 12. 0.86125 0.81632 0.71528] [ 13. 0.85973 0.81679 0.71329] [ 14. 0.96085 0.85626 0.69921] [ 15. 0.93619 0.83843 0.70179] [ 16. 0.92995 0.83946 0.70367] [ 17. 0.90384 0.82234 0.71518] [ 18. 0.89965 0.81864 0.70972] [ 19. 0.88279 0.82276 0.70913] [ 20. 0.87067 0.81281 0.7125 ] [ 21. 0.84173 0.81346 0.71369] [ 22. 0.82026 0.8091 0.71677] [ 23. 0.80922 0.80768 0.71944] [ 24. 0.79314 0.79963 0.72133] [ 25. 0.78475 0.79911 0.72202] [ 26. 0.77612 0.79913 0.72093] [ 27. 0.76019 0.79675 0.7244 ] [ 28. 0.74892 0.79658 0.72361] [ 29. 0.75726 0.79311 0.72411]
CPU times: user 23min 18s, sys: 6min 55s, total: 30min 14s Wall time: 20min 22s
fig = plt.figure(figsize=(15, 10))
plt.plot(loss_adamw1)
plt.plot(loss_adamw2)
plt.plot(loss_adamw3)
plt.show()
%%time
# With weight decay
adamw = Get_Adam()
lr = [1e-5,1e-4,1e-3]
wd = [1e-5,1e-4,1e-3]
loss_adamw4, learn4 = experiment(adamw, PATH, lr, use_wd_sched=True, wds=wd, do_unfreeze=True)
save_list('adamw_loss4.txt', loss_adamw4)
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.89021 1.49814 0.50179] Now with precompute as False
Failed to display Jupyter Widget of type HBox.
If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean that the widgets JavaScript is still loading. If this message persists, it likely means that the widgets JavaScript library is either not installed or not enabled. See the Jupyter Widgets Documentation for setup instructions.
If you're reading this message in another frontend (for example, a static rendering on GitHub or NBViewer), it may mean that your frontend doesn't currently support widgets.
[ 0. 1.14415 0.98321 0.65486] [ 1. 1.17225 1.1191 0.66458] [ 2. 1.07665 0.9536 0.66597] [ 3. 1.03984 0.93191 0.68105] [ 4. 1.02686 0.9367 0.68978] [ 5. 1.19762 1.20376 0.69494] [ 6. 1.09923 0.97489 0.65992] [ 7. 1.07333 0.95225 0.66568] [ 8. 1.05902 0.93139 0.67877] [ 9. 1.05099 0.92172 0.68036] [ 10. 1.02277 0.95366 0.68105] [ 11. 1.02635 0.95059 0.68829] [ 12. 1.10198 1.06264 0.69732] [ 13. 1.54751 1.5987 0.69246] [ 14. 1.11582 1.00986 0.64722] [ 15. 1.10576 0.97402 0.662 ] [ 16. 1.08318 0.96627 0.66181] [ 17. 1.08801 0.96633 0.66399] [ 18. 1.07158 0.95688 0.66409] [ 19. 1.06059 0.94413 0.66726] [ 20. 1.05009 0.92797 0.67669] [ 21. 1.05665 0.96412 0.66677] [ 22. 1.06465 0.94497 0.67788] [ 23. 1.03009 0.938 0.68066] [ 24. 1.04048 0.94522 0.68006] [ 25. 1.04625 0.97119 0.68502] [ 26. 1.06764 1.00826 0.68929] [ 27. 1.21396 1.16204 0.68234] [ 28. 1.55344 1.5575 0.65724] [ 29. 2.04397 2.08168 0.64682]
CPU times: user 23min 22s, sys: 6min 57s, total: 30min 19s Wall time: 20min 27s