notebook.community

Edit and run



In [1]:

    
%reload_ext autoreload
%autoreload 2
%matplotlib inline

from fastai.imports import *
from fastai.conv_learner import *
from fastai.torch_imports import *
from fastai.model import *
from fastai.dataset import *
from fastai.sgdr import *
from fastai.plots import *



In [2]:

    
PATH = "data/dogbreeds/"



In [3]:

    
arch = resnext50
sz = 224
bs = 64



In [4]:

    
labels_csv = f'{PATH}labels.csv'
n = len(list(open(labels_csv)))-1
val_idxs = get_cv_idxs(n)



In [1]:

    
def get_data(sz, bs):
    tfms = tfms_from_model(arch, sz, aug_tfms=transforms_side_on, max_zoom=1.1)
    data = ImageClassifierData.from_csv(PATH, 'train', labels_csv, bs=bs, tfms=tfms, 
                                        val_idxs=val_idxs, suffix='.jpg', test_name='test')
    return data if sz > 300 else data.resize(340, 'tmp')
# see: http://forums.fast.ai/t/dog-breed-identification-challenge/7464/51

In the below code, why are we choosing 300 as the default size value to check if condition?

<><><><>

Great question. Since we have max_zoom=1.1, I figured we should ensure our images are at release sz*1.1. And I figured resizing them to 340x340 would save plenty of time, and leave plenty of room to experiment.

http://forums.fast.ai/t/dog-breed-identification-challenge/7464/51

Note this notebook was run with ... if sz < 300 ... since I didn't understand what was going on.



In [40]:

    
# data = get_data(sz, bs)

# labels_df = pd.read_csv(labels_csv)
# labels_df.pivot_table(index='breed', aggfunc=len).sort_values('id', ascending=False)
# fn = PATH + data.trn_ds.fnames[0]
# PIL.Image.open(fn)

# size_d = {k: PIL.Image.open(PATH+k).size for k in data.trn_ds.fnames}



In [41]:

    
# row_sz, col_sz = list(zip(*size_d.values()))
# row_sz, col_sz = np.array(row_sz), np.array(col_sz)

# plt.hist(row_sz[row_sz < 1000]); plt.hist(col_sz[col_sz < 1000]);



In [6]:

    
from sklearn import metrics



In [44]:

    
data = get_data(sz, bs)
learn = ConvLearner.pretrained(arch, data, precompute=True)









    



100%|██████████| 128/128 [02:53<00:00,  1.36s/it]
100%|██████████| 32/32 [00:43<00:00,  1.37s/it]
100%|██████████| 162/162 [03:39<00:00,  1.35s/it]



In [49]:

    
learn.lr_find()
learn.sched.plot()









    





 
 










    



 93%|█████████▎| 119/128 [00:01<00:00, 114.50it/s, loss=12.1]



In [50]:

    
learn.fit(2e-2, 2)









    





 
 










    



[ 0.       0.85798  0.39835  0.88994]                         
[ 1.       0.42135  0.31747  0.90664]



In [51]:

    
learn.fit(2e-2, 2)









    





 
 










    



[ 0.       0.27853  0.30884  0.90023]                         
[ 1.       0.21362  0.29688  0.90758]



In [55]:

    
learn.precompute=False
learn.fit(2e-2, 2)









    





 
 










    



[ 0.       0.31715  0.30835  0.89971]                        
[ 1.       0.27991  0.29524  0.90947]



In [56]:

    
learn.save('RNx50_224_pre')



In [11]:

    
# increasing size - taking advtg of Fully-Convolutional Arch
learn.set_data(get_data(299, 48))



In [65]:

    
learn.fit(1e-2, 3, cycle_len=1)









    





 
 










    



[ 0.       0.26006  0.24794  0.92193]                        
[ 1.       0.23345  0.23904  0.92906]                        
[ 2.       0.21016  0.23497  0.93182]



In [69]:

    
learn.save('RNx50_224_pre')



In [8]:

    
data  = get_data(299, 48)
learn = ConvLearner.pretrained(arch, data)
learn.load('RNx50_224_pre')



In [10]:

    
learn.freeze()



In [12]:

    
log_preds, y = learn.TTA()
probs = np.exp(log_preds)
accuracy(log_preds, y), metrics.log_loss(y, probs)









    



  0%|          | 0/4 [00:00<?, ?it/s]
 25%|██▌       | 1/4 [01:29<04:28, 89.43s/it]
 50%|█████     | 2/4 [02:58<02:58, 89.47s/it]
 75%|███████▌  | 3/4 [04:28<01:29, 89.42s/it]
100%|██████████| 4/4 [05:57<00:00, 89.41s/it]
                                             





    Out[12]:





(0.92906066536203524, 0.22787421921797549)



In [13]:

    
test_preds = np.exp(learn.TTA(is_test=True)[0])









    



  0%|          | 0/4 [00:00<?, ?it/s]
 25%|██▌       | 1/4 [07:20<22:02, 440.87s/it]
 50%|█████     | 2/4 [14:41<14:41, 440.92s/it]
 75%|███████▌  | 3/4 [22:02<07:20, 440.86s/it]
100%|██████████| 4/4 [29:23<00:00, 440.87s/it]



In [ ]:



In [ ]:



In [ ]:

Rerunning without validation for predictions:



In [ ]:

    
from sklearn import metrics
PATH = "data/dogbreeds/"
arch = resnext50
sz = 224
bs = 64
labels_csv = f'{PATH}labels.csv'
# n = len(list(open(labels_csv)))-1
val_idxs = get_cv_idxs(0)
def get_data(sz, bs):
    tfms = tfms_from_model(arch, sz, aug_tfms=transforms_side_on, max_zoom=1.1)
    data = ImageClassifierData.from_csv(PATH, 'train', labels_csv, bs=bs, tfms=tfms, 
                                        val_idxs=val_idxs, suffix='.jpg', test_name='test')
    return data if sz < 300 else data.resize(340, 'tmp')



In [ ]:

    
data = get_data(sz, bs)
learn = ConvLearner.pretrained(arch, data, precompute=True)
learn.fit(1e-2, 2)
learn.precompute=False
learn.fit(1e-2, 5, cycle_len=1)
learn.set_data(get_data(299, bs=32))
learn.fit(1e-2, 3, cycle_len=1)
learn.fit(1e-2, 3, cycle_len=1, cycle_mult=2)