In [2]:

%reload_ext autoreload
%autoreload 2

from fastai import *
from fastai.vision import *
from fastai.docs import *

Dogs and cats¶

Resnet 34¶

In [3]:

untar_data(DOGS_PATH)

HBox(children=(IntProgress(value=0, max=839285364), HTML(value='')))

In [11]:

data = image_data_from_folder(DOGS_PATH, ds_tfms=get_transforms(), tfms=imagenet_norm, size=224)
img,label = data.valid_ds[-1]
print(data.classes[label])
img

dogs

Out[11]:

In [12]:

learn = ConvLearner(data, tvm.resnet34, metrics=accuracy)
learn.fit_one_cycle(1)

VBox(children=(HBox(children=(IntProgress(value=0, max=1), HTML(value='0.00% [0/1 00:00<00:00]'))), HTML(value…

Total time: 00:49
epoch  train loss  valid loss  accuracy
0      0.058669    0.028334    0.991500  (00:49)

In [14]:

learn.unfreeze()
learn.fit_one_cycle(6, slice(1e-5,3e-4), pct_start=0.05)

VBox(children=(HBox(children=(IntProgress(value=0, max=6), HTML(value='0.00% [0/6 00:00<00:00]'))), HTML(value…

Total time: 06:40
epoch  train loss  valid loss  accuracy
0      0.045248    0.031617    0.986000  (01:05)
1      0.031750    0.021102    0.992000  (01:06)
2      0.016185    0.021817    0.993000  (01:06)
3      0.015322    0.023137    0.991500  (01:07)
4      0.009295    0.021500    0.994000  (01:07)
5      0.008227    0.020382    0.994000  (01:07)

In [15]:

accuracy(*learn.TTA())

HBox(children=(IntProgress(value=0, max=16), HTML(value='')))

HBox(children=(IntProgress(value=0, max=16), HTML(value='0.00% [0/16 00:00<00:00]')))

VBox(children=(HBox(children=(IntProgress(value=0, max=8), HTML(value='0.00% [0/8 00:00<00:00]'))), HTML(value…

Total time: 00:33

Out[15]:

tensor(0.9965)

rn50¶

In [16]:

learn = ConvLearner(data, tvm.resnet50, metrics=accuracy)
learn.fit_one_cycle(6)

VBox(children=(HBox(children=(IntProgress(value=0, max=6), HTML(value='0.00% [0/6 00:00<00:00]'))), HTML(value…

Total time: 09:24
epoch  train loss  valid loss  accuracy
0      0.054099    0.024692    0.992500  (01:36)
1      0.039004    0.025797    0.991500  (01:33)
2      0.040528    0.021088    0.992000  (01:33)
3      0.026295    0.022694    0.993000  (01:33)
4      0.019064    0.016771    0.994000  (01:33)
5      0.016541    0.015919    0.995500  (01:33)

In [17]:

learn.unfreeze()
learn.fit_one_cycle(6, slice(1e-5,3e-4), pct_start=0.05)

VBox(children=(HBox(children=(IntProgress(value=0, max=6), HTML(value='0.00% [0/6 00:00<00:00]'))), HTML(value…

Total time: 12:42
epoch  train loss  valid loss  accuracy
0      0.031718    0.019076    0.991000  (02:07)
1      0.031773    0.025698    0.991000  (02:07)
2      0.014313    0.032038    0.990500  (02:06)
3      0.009425    0.024969    0.991500  (02:06)
4      0.006091    0.016411    0.995500  (02:06)
5      0.004848    0.016143    0.995500  (02:07)

In [18]:

accuracy(*learn.TTA())

HBox(children=(IntProgress(value=0, max=16), HTML(value='')))

HBox(children=(IntProgress(value=0, max=16), HTML(value='0.00% [0/16 00:00<00:00]')))

VBox(children=(HBox(children=(IntProgress(value=0, max=8), HTML(value='0.00% [0/8 00:00<00:00]'))), HTML(value…

Total time: 00:45

Out[18]:

tensor(0.9965)