Imagenette/woof training
This is an example of how to analyze the results logged by a number of runs of fastgpu on Imagenette.
#all_slow
#hide
from fastcore.all import *
from nbdev.showdoc import *
from pandas import DataFrame
from IPython.lib.display import FileLink
path = Path('~/git/fastai/nbs/examples/exps').expanduser()
path_out = path/'out'
path_scr = path/'complete'
def fn_name(fn): return fn.name.split('.')[0]
hdrs = 'name epoch trn_loss val_loss top1 top5 time'.split()
def scr_results(fn):
return L([dict(zip(hdrs, [fn_name(fn)] + o.strip().split()))
for o in fn.readlines() if re.match(r'\d+ ', o)])
all_results = L([scr_results(o) for o in path_out.glob('*.stdout')]).concat()
df_results = DataFrame.from_records(all_results).query('time==time').astype(
dict(epoch=int, trn_loss=float, val_loss=float, top5=float, top1=float))
def scr_params(fn):
return merge(dict([o.strip().split() for o in fn.read_text().split('--')[1:]]),
{'name':fn_name(fn)})
df_params = DataFrame([scr_params(o) for o in path_scr.ls()]).astype({'epochs':int})
df_params.head()
| woof | lr | size | sqrmom | mom | eps | epochs | bs | opt | sa | fp16 | arch | name | pool | meta | runs | mixup | sh | sym | beta | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | 0 | 8e-3 | 192 | 0.99 | 0.95 | 1e-6 | 5 | 64 | ranger | 1 | 0 | xse_resnext18 | 02a | NaN | NaN | NaN | NaN | NaN | NaN | NaN |
| 1 | 0 | 8e-3 | 128 | 0.99 | 0.95 | 1e-6 | 20 | 64 | ranger | 1 | 1 | xse_resnext50 | xf104a | MaxPool | NaN | NaN | NaN | NaN | NaN | NaN |
| 2 | 0 | 8e-3 | 256 | 0.99 | 0.95 | 1e-6 | 5 | 64 | ranger | 1 | 0 | xse_resnext18 | 03a | NaN | NaN | NaN | NaN | NaN | NaN | NaN |
| 3 | 1 | 8e-3 | 128 | 0.99 | 0.95 | 1e-6 | 5 | 64 | ranger | 1 | 1 | xse_resnext50 | x01b | NaN | NaN | NaN | NaN | NaN | NaN | NaN |
| 4 | 0 | 8e-3 | 192 | 0.99 | 0.95 | 1e-6 | 20 | 64 | ranger | 1 | 1 | xse_resnext50 | x05a | NaN | NaN | NaN | NaN | NaN | NaN | NaN |
df_merge = pd.merge(df_params, df_results, 'left', on='name')
assert df_merge[pd.isna(df_merge.arch)].empty
df_merge.fillna(0.0, inplace=True)
df_merge['repoch'] = df_merge.epochs - df_merge.epoch
df_merge.to_csv('imagenette.csv', index=False)
FileLink('imagenette.csv')