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
- leaky relu / elu
Planet Kaggle competition¶
In [ ]:
%reload_ext autoreload
%autoreload 2
%matplotlib inline
In [ ]:
from fast_gen import *
from learner import *
from pt_models import *
from dataset_pt import *
from sgdr_pt import *
from planet import *
bs=64; f_model = resnet34
path = "/data/jhoward/fast/planet/"
torch.cuda.set_device(1)
In [ ]:
n=len(list(open(f'{path}train_v2.csv')))-1
In [ ]:
data=get_data_pad(f_model, path, 256, 64, n, 0)
In [ ]:
learn = Learner.pretrained_convnet(f_model, data, metrics=[f2])
Train¶
In [ ]:
learn.fit(0.2, 1, cycle_len=1)
[0.15438985784802958, 0.13450478, 0.87891029153421008]
In [ ]:
learn.sched.plot_lr()
In [ ]:
learn.unfreeze()
In [ ]:
learn.fit([0.01,0.05,0.2], 12, cycle_len=4)
[0.11952567681830356, 0.10556465, 0.90650807074212403]
[0.11009075188366914, 0.096736901, 0.91672481730802124]
[0.10827947736338327, 0.093592748, 0.91964642671839225]
[0.1058942063357304, 0.092603102, 0.92130033558366631]
[0.1092922736445057, 0.095529161, 0.91819015894496181]
[0.1011713251614138, 0.092735209, 0.92159361463751932]
[0.09583996319032348, 0.09039887, 0.92251825740384097]
[0.09977698443701977, 0.089664221, 0.92254781482418324]
[0.10362990970969754, 0.092073925, 0.92141614562971053]
[0.10036470544572228, 0.090630345, 0.92154762597476325]
[0.09469915216101382, 0.089381441, 0.92234579315867105]
[0.0912562884700505, 0.089002736, 0.92252729549243617]
In [ ]:
learn.fit([1e-4,1e-3,0.01], 4)
[0.09916652571222209, 0.090306021, 0.9227976239444905]
[0.09459107546738892, 0.089821696, 0.92330719115188997]
[0.0965355865038755, 0.089525335, 0.92353226343888395]
[0.09088765592814434, 0.089346372, 0.92387145358147427]
Evaluate¶
In [ ]:
name = '170809'
In [ ]:
def load_cycle_cv(cv, cycle):
data=get_data_zoom(f_model, path, 256, 64, n, cv)
learn.set_data(data)
learn.load_cycle(f'{name}_{cv}', cycle)
return data
In [ ]:
data = load_cycle_cv(0,1)
In [ ]:
val = learn.predict()
In [ ]:
f2(val,data.val_y)
Out[ ]:
0.92719935814702104
In [ ]:
f2(learn.TTA(),data.val_y)
Out[ ]:
0.92776745163594243
In [ ]:
f2(val,data.val_y)
Out[ ]:
0.93020923622107132
In [ ]:
f2(learn.TTA(),data.val_y)
Out[ ]:
0.93146299154229428
In [ ]:
def get_labels(a): return [data.classes[o] for o in a.nonzero()[0]]
lbls = test>0.2
idx=9
print(get_labels(lbls[idx]))
PIL.Image.open(path+data.test_dl.dataset.fnames[idx]).convert('RGB')
['partly_cloudy', 'primary']
Out[ ]:
In [ ]:
res = [get_labels(o) for o in lbls]
data.test_dl.dataset.fnames[:5]
Out[ ]:
['test-jpg/test_17555.jpg', 'test-jpg/test_27899.jpg', 'test-jpg/test_21010.jpg', 'test-jpg/test_39610.jpg', 'test-jpg/test_34768.jpg']
In [ ]:
outp = pd.DataFrame({'image_name': [f[9:-4] for f in data.test_dl.dataset.fnames],
'tags': [' '.join(l) for l in res]})
outp.head()
Out[ ]:
| image_name | tags | |
|---|---|---|
| 0 | test_17555 | agriculture haze primary road water |
| 1 | test_27899 | agriculture clear partly_cloudy primary road w... |
| 2 | test_21010 | agriculture bare_ground clear primary road |
| 3 | test_39610 | agriculture clear primary road |
| 4 | test_34768 | agriculture clear primary road |
In [ ]:
outp.to_csv('tmp/subm.gz', compression='gzip', index=None)
In [ ]:
from IPython.display import FileLink
In [ ]:
FileLink('tmp/subm.gz')
Out[ ]:
In [ ]:
def cycle_preds(name, cycle, n_tta=4, is_test=False):
learn.load_cycle(name, cycle)
return learn.TTA(n_tta, is_test=is_test)
def cycle_cv_preds(cv, n_tta=4, is_test=False):
data=get_data_pad(f_model, path, 256, 64, n, cv)
learn.set_data(data)
return [cycle_preds(f'{name}_{cv}',i, is_test=is_test) for i in range(5)]
- check dogs and cats
- get resize working again with new path structure
In [ ]:
%%time
preds_arr=[]
for i in range(5):
print(i)
preds_arr.append(cycle_cv_preds(i, is_test=True))
0 1 2 3 4 CPU times: user 6h 23min 45s, sys: 16min 59s, total: 6h 40min 44s Wall time: 6h 43min 38s
In [ ]:
def all_cycle_cv_preds(end_cycle, start_cycle=0, n_tta=4, is_test=False):
return [cycle_cv_preds(i, is_test=is_test) for i in range(start_cycle, end_cycle)]
In [ ]:
np.savez_compressed(f'{path}tmp/test_preds', preds_arr)
In [ ]:
preds_avg = [np.mean(o,0) for o in preds_arr]
test = np.mean(preds_avg,0)
In [ ]:
%time preds_arr = all_cycle_cv_preds(5)
CPU times: user 54min 11s, sys: 2min 40s, total: 56min 52s Wall time: 56min 18s
In [ ]:
[f2(preds_arr[0][o],data.val_y) for o in range(5)]
Out[ ]:
[0.93071536605914396, 0.93116843932677196, 0.93069579784787815, 0.9307239826054069, 0.93051232753450475]
In [ ]:
preds_avg = [np.mean(o,0) for o in preds_arr]
In [ ]:
ys = [get_data_zoom(f_model, path, 256, 64, n, cv).val_y for cv in range(5)]
In [ ]:
f2s = [f2(o,y) for o,y in zip(preds_avg,ys)]; f2s
Out[ ]:
[0.93137364725177285, 0.93053278668158479, 0.93297669347979839, 0.9323465548921448, 0.93621843040962094]
In [ ]:
ots = [opt_th(o,y) for o,y in zip(preds_avg,ys)]; ots
Out[ ]:
[0.20000000000000004, 0.17000000000000001, 0.21000000000000005, 0.23000000000000007, 0.19000000000000003]
In [ ]:
np.mean(ots)
Out[ ]:
0.20000000000000004
In [ ]:
np.mean(f2s,0)
Out[ ]:
0.93268962254298438