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

Fisheries competition¶

In this notebook we're going to investigate a range of different techniques for the Kaggle fisheries competition. In this competition, The Nature Conservancy asks you to help them detect which species of fish appears on a fishing boat, based on images captured from boat cameras of various angles. Your goal is to predict the likelihood of fish species in each picture. Eight target categories are available in this dataset: Albacore tuna, Bigeye tuna, Yellowfin tuna, Mahi Mahi, Opah, Sharks, Other

You can use this api to download the data from Kaggle.

In [ ]:

# Put these at the top of every notebook, to get automatic reloading and inline plotting
%reload_ext autoreload
%autoreload 2
%matplotlib inline

# This file contains all the main external libs we'll use
from fastai.imports import *
from fastai.plots import *
from fastai.io import get_data

PATH = "data/fish/"

First look at fish pictures¶

In [ ]:

!ls {PATH}

In [ ]:

!ls {PATH}train

In [ ]:

files = !ls {PATH}train/ALB | head
files

In [ ]:

img = plt.imread(f'{PATH}train/ALB/{files[0]}')
plt.imshow(img);

Data pre-processing¶

Here we are changing the structure of the training data to make it more convinient. We will have all images in a common directory images and will have a file train.csv with all labels.

In [ ]:

from os import listdir
from os.path import join
train_path = f'{PATH}/train'

In [ ]:

dirs = [d for d in listdir(train_path) if os.path.isdir(join(train_path,d))]
print(dirs)

In [ ]:

train_dict = {d: listdir(join(train_path, d)) for d in dirs}

In [ ]:

train_dict["LAG"][:10]

In [ ]:

sum(len(v) for v in train_dict.values())

In [ ]:

with open(f"{PATH}train.csv", "w") as csv:
    csv.write("img,label\n")
    for d in dirs:
        for f in train_dict[d]: csv.write(f'{f},{d}\n')

In [ ]:

img_path = f'{PATH}images'
os.makedirs(img_path, exist_ok=True)

In [ ]:

!cp {PATH}train/*/*.jpg {PATH}images/

Our first model with Center Cropping¶

Here we import the libraries we need. We'll learn about what each does during the course.

In [ ]:

from fastai.transforms import *
from fastai.conv_learner import *
from fastai.model import *
from fastai.dataset import *
from fastai.sgdr import *

In [ ]:

sz=350
bs=64
csv_fname = os.path.join(PATH, "train.csv")
train_labels = list(open(csv_fname))
n = len(list(open(csv_fname)))-1
val_idxs = get_cv_idxs(n)

In [ ]:

tfms = tfms_from_model(resnet34, sz)
data = ImageClassifierData.from_csv(PATH, "images", csv_fname, bs, tfms, val_idxs)

In [ ]:

learn = ConvLearner.pretrained(resnet34, data, precompute=True, opt_fn=optim.Adam, ps=0.5)

models are fine
supper init
  0%|          | 0/48 [00:00<?, ?it/s]torch.Size([64, 3, 350, 350])

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-5-2599e6e01994> in <module>()
----> 1 learn = ConvLearner.pretrained(resnet34, data, precompute=True, opt_fn=optim.Adam, ps=0.5)

~/github/fastai/courses/dl1/fastai/conv_learner.py in pretrained(self, f, data, ps, xtra_fc, xtra_cut, **kwargs)
     89         models = ConvnetBuilder(f, data.c, data.is_multi, data.is_reg, ps=ps, xtra_fc=xtra_fc, xtra_cut=xtra_cut)
     90         print("models are fine")
---> 91         return self(data, models, **kwargs)
     92 
     93     @property

~/github/fastai/courses/dl1/fastai/conv_learner.py in __init__(self, data, models, precompute, **kwargs)
     81         elif self.metrics is None:
     82             self.metrics = [accuracy_multi] if self.data.is_multi else [accuracy]
---> 83         self.save_fc1()
     84         self.freeze()
     85         self.precompute=precompute

~/github/fastai/courses/dl1/fastai/conv_learner.py in save_fc1(self)
    123         if len(self.activations[0])==0:
    124             m=self.models.top_model
--> 125             predict_to_bcolz(m, self.data.fix_dl, act)
    126             predict_to_bcolz(m, self.data.val_dl, val_act)
    127             if self.data.test_dl: predict_to_bcolz(m, self.data.test_dl, test_act)

~/github/fastai/courses/dl1/fastai/model.py in predict_to_bcolz(m, gen, arr, workers)
     11     for x,*_ in tqdm(gen):
     12         print(x.shape)
---> 13         y = to_np(m(VV(x)).data)
     14         with lock:
     15             arr.append(y)

~/github/fastai/courses/dl1/fastai/core.py in VV(x)
     16 def V(x):   return [V_(o) for o in x] if isinstance(x,list) else V_(x)
     17 def VV_(x): return x.cuda(async=True) if isinstance(x, Variable) else Variable(x.cuda(async=True), volatile=True)
---> 18 def VV(x):  return [VV_(o) for o in x] if isinstance(x,list) else VV_(x)
     19 
     20 def to_np(v):

~/github/fastai/courses/dl1/fastai/core.py in VV_(x)
     15 def V_(x):  return x.cuda(async=True) if isinstance(x, Variable) else Variable(x.cuda(async=True))
     16 def V(x):   return [V_(o) for o in x] if isinstance(x,list) else V_(x)
---> 17 def VV_(x): return x.cuda(async=True) if isinstance(x, Variable) else Variable(x.cuda(async=True), volatile=True)
     18 def VV(x):  return [VV_(o) for o in x] if isinstance(x,list) else VV_(x)
     19 

~/anaconda/lib/python3.6/site-packages/torch/_utils.py in _cuda(self, device, async)
     64         else:
     65             new_type = getattr(torch.cuda, self.__class__.__name__)
---> 66             return new_type(self.size()).copy_(self, async)
     67 
     68 

RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1502009910772/work/torch/lib/THC/generic/THCStorage.cu:66

In [ ]:

lrf=learn.lr_find()
learn.sched.plot()

 65%|██████▍   | 31/48 [00:00<00:03,  4.59it/s, loss=2.07]

In [ ]:

learn.fit(0.01, 4, cycle_len=1, cycle_mult=2)

[ 0.      0.6912  0.617   0.7949]                         
[ 1.      0.6748  0.3572  0.8928]                          
[ 2.      0.5156  0.3283  0.8974]                          
[ 3.      0.4837  0.3402  0.8873]                          
[ 4.      0.4303  0.2782  0.9179]                          
[ 5.      0.3515  0.2209  0.9267]                          
[ 6.      0.2841  0.2082  0.9349]                          
[ 7.      0.318   0.2482  0.9287]                          
[ 8.      0.3518  0.2316  0.9323]                          
[ 9.      0.3343  0.2133  0.9323]                         
[ 10.       0.2857   0.2128   0.9293]                      
[ 11.       0.2414   0.2148   0.9349]                      
[ 12.       0.2077   0.189    0.9287]                      
[ 13.       0.1834   0.1826   0.9352]                      
[ 14.       0.162    0.179    0.9391]

In [ ]:

lrs=np.array([1e-4,1e-3,1e-2])
learn.precompute=False

learn.freeze_to(6)
lrf=learn.lr_find(lrs/1e3)
learn.sched.plot()

 79%|███████▉  | 38/48 [00:24<00:05,  1.89it/s, loss=1.48]

Same model with No cropping¶

NOTE: Before running this remove the temp file under data/fish.

In [ ]:

sz = 350
tfms = tfms_from_model(resnet34, sz, crop_type=CropType.NO)
data = ImageClassifierData.from_csv(PATH, "images", csv_fname, bs, tfms, val_idxs)

In [ ]:

learn = ConvLearner.pretrained(resnet34, data, precompute=True, opt_fn=optim.Adam, ps=0.5)

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-8-2599e6e01994> in <module>()
----> 1 learn = ConvLearner.pretrained(resnet34, data, precompute=True, opt_fn=optim.Adam, ps=0.5)

~/github/fastai/courses/dl1/fastai/conv_learner.py in pretrained(self, f, data, ps, xtra_fc, xtra_cut, **kwargs)
     87     @classmethod
     88     def pretrained(self, f, data, ps=None, xtra_fc=None, xtra_cut=0, **kwargs):
---> 89         models = ConvnetBuilder(f, data.c, data.is_multi, data.is_reg, ps=ps, xtra_fc=xtra_fc, xtra_cut=xtra_cut)
     90         print("models are fine")
     91         return self(data, models, **kwargs)

~/github/fastai/courses/dl1/fastai/conv_learner.py in __init__(self, f, c, is_multi, is_reg, ps, xtra_fc, xtra_cut)
     40         fc_layers = self.get_fc_layers()
     41         self.n_fc = len(fc_layers)
---> 42         self.fc_model = nn.Sequential(*fc_layers).cuda()
     43         apply_init(self.fc_model, kaiming_normal)
     44         self.model=nn.Sequential(*(layers+fc_layers)).cuda()

~/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in cuda(self, device_id)
    145                 copied to that device
    146         """
--> 147         return self._apply(lambda t: t.cuda(device_id))
    148 
    149     def cpu(self):

~/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in _apply(self, fn)
    116     def _apply(self, fn):
    117         for module in self.children():
--> 118             module._apply(fn)
    119 
    120         for param in self._parameters.values():

~/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in _apply(self, fn)
    122                 # Variables stored in modules are graph leaves, and we don't
    123                 # want to create copy nodes, so we have to unpack the data.
--> 124                 param.data = fn(param.data)
    125                 if param._grad is not None:
    126                     param._grad.data = fn(param._grad.data)

~/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py in <lambda>(t)
    145                 copied to that device
    146         """
--> 147         return self._apply(lambda t: t.cuda(device_id))
    148 
    149     def cpu(self):

~/anaconda/lib/python3.6/site-packages/torch/_utils.py in _cuda(self, device, async)
     64         else:
     65             new_type = getattr(torch.cuda, self.__class__.__name__)
---> 66             return new_type(self.size()).copy_(self, async)
     67 
     68 

RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1502009910772/work/torch/lib/THC/generic/THCStorage.cu:66

In [ ]:

lrf=learn.lr_find()
learn.sched.plot()

 56%|█████▋    | 27/48 [00:00<00:02, 10.40it/s, loss=2.16]

In [ ]:

learn.fit(0.01, 4, cycle_len=1, cycle_mult=2)

[ 0.      0.682   0.5775  0.8251]                         
[ 1.      0.6791  0.3741  0.8824]                          
[ 2.      0.5242  0.3172  0.9056]                          
[ 3.      0.4669  0.2973  0.8942]                         
[ 4.      0.4197  0.2335  0.9238]                          
[ 5.      0.3512  0.2161  0.9277]                          
[ 6.      0.2884  0.181   0.9306]                          
[ 7.      0.3255  0.2539  0.9176]                          
[ 8.      0.3452  0.2123  0.9284]                          
[ 9.      0.317   0.2007  0.9371]                          
[ 10.       0.2644   0.1644   0.9557]                      
[ 11.       0.2244   0.1454   0.9528]                      
[ 12.       0.1818   0.1343   0.9665]                      
[ 13.       0.1599   0.1404   0.9599]                      
[ 14.       0.1546   0.1408   0.9586]

In [ ]:

lrs=np.array([1e-4,1e-3,1e-2])
learn.precompute=False

learn.unfreeze()
lrf=learn.lr_find(lrs/1e3)
learn.sched.plot()

 79%|███████▉  | 38/48 [00:36<00:07,  1.25it/s, loss=1.65]

In [ ]:

lrs=np.array([1e-5,1e-4,1e-3])
learn.fit(lrs, 5, cycle_len=1, cycle_mult=2)

                                                          
                                                          
                                                          
                                                          

                                                          
                                                          

                                                          

  2%|▏         | 1/48 [00:05<04:30,  5.75s/it, loss=0.74]
  4%|▍         | 2/48 [00:06<03:16,  4.26s/it, loss=0.936]
  6%|▋         | 3/48 [00:07<02:25,  3.23s/it, loss=0.874]
  8%|▊         | 4/48 [00:08<01:49,  2.50s/it, loss=0.754]
[ 0.      0.2649  0.1643  0.9404]                          
[ 1.      0.1998  0.1073  0.9681]                          
[ 2.      0.1288  0.1007  0.9694]                          
[ 3.      0.0923  0.1089  0.9736]                           
[ 4.      0.0843  0.1173  0.971 ]                           
[ 5.      0.0549  0.1002  0.9694]                           
[ 6.      0.0396  0.0964  0.9733]                           
[ 7.      0.0315  0.1249  0.9762]                           
[ 8.      0.0382  0.1132  0.9746]                           
[ 9.      0.0406  0.1013  0.973 ]                           
[ 10.       0.0285   0.1085   0.9749]                       
[ 11.       0.0202   0.1131   0.9762]                       
[ 12.       0.0154   0.1048   0.9772]                       
[ 13.       0.0308   0.0956   0.9785]                        
[ 14.       0.0166   0.1028   0.9788]                       
[  1.5000e+01   1.1636e-02   1.0597e-01   9.7753e-01]       
[  1.6000e+01   1.1961e-02   1.0537e-01   9.7623e-01]       
[  1.7000e+01   1.2807e-02   1.0589e-01   9.7753e-01]        
[  1.8000e+01   1.4130e-02   1.3264e-01   9.7493e-01]       
[  1.9000e+01   1.6299e-02   1.1089e-01   9.8014e-01]       
[  2.0000e+01   1.6554e-02   1.3690e-01   9.7493e-01]       
[  2.1000e+01   1.4021e-02   1.5582e-01   9.6809e-01]       
[  2.2000e+01   9.8160e-03   1.3994e-01   9.7363e-01]        
[  2.3000e+01   7.9080e-03   1.3732e-01   9.7590e-01]        
[  2.4000e+01   1.7213e-02   1.5735e-01   9.7330e-01]        
[  2.5000e+01   2.0878e-02   1.5379e-01   9.7363e-01]       
[  2.6000e+01   2.0992e-02   1.5257e-01   9.7363e-01]       
[  2.7000e+01   1.8790e-02   1.5161e-01   9.7330e-01]       
[  2.8000e+01   1.1324e-02   1.4440e-01   9.7199e-01]       
[  2.9000e+01   7.9810e-03   1.4998e-01   9.7363e-01]        
[  3.0000e+01   6.6660e-03   1.4507e-01   9.7493e-01]

Predicting bounding boxes¶

Getting bounding boxes data¶

This part needs to run just the first time to get the file trn_bb_labels

In [ ]:

import json
anno_classes = ['alb', 'bet', 'dol', 'lag', 'other', 'shark', 'yft']

In [ ]:

def get_annotations():
    annot_urls = [
        '5458/bet_labels.json', '5459/shark_labels.json', '5460/dol_labels.json',
        '5461/yft_labels.json', '5462/alb_labels.json', '5463/lag_labels.json'
    ]
    cache_subdir = os.path.abspath(os.path.join(PATH, 'annos'))
    url_prefix = 'https://kaggle2.blob.core.windows.net/forum-message-attachments/147157/'
    os.makedirs(cache_subdir, exist_ok=True)
    
    for url_suffix in annot_urls:
        fname = url_suffix.rsplit('/', 1)[-1]
        get_data(url_prefix + url_suffix, f'{cache_subdir}/{fname}')

In [ ]:

# run this code to get annotation files
get_annotations()

In [ ]:

# creates a dictionary of all annotations per file
bb_json = {}
for c in anno_classes:
    if c == 'other': continue # no annotation file for "other" class
    j = json.load(open(f'{PATH}annos/{c}_labels.json', 'r'))
    for l in j:
        if 'annotations' in l.keys() and len(l['annotations'])>0:
            bb_json[l['filename'].split('/')[-1]] = sorted(
                l['annotations'], key=lambda x: x['height']*x['width'])[-1]
bb_json['img_04908.jpg']

In [ ]:

raw_filenames = pd.read_csv(csv_fname)["img"].values

In [ ]:

file2idx = {o:i for i,o in enumerate(raw_filenames)}

empty_bbox = {'height': 0., 'width': 0., 'x': 0., 'y': 0.}
for f in raw_filenames:
    if not f in bb_json.keys(): bb_json[f] = empty_bbox

bb_params = ['height', 'width', 'x', 'y']
def convert_bb(bb):
    bb = [bb[p] for p in bb_params]
    bb[2] = max(bb[2], 0)
    bb[3] = max(bb[3], 0)
    return bb

In [ ]:

trn_bbox = np.stack([convert_bb(bb_json[f]) for f in raw_filenames]).astype(np.float32)
trn_bb_labels = [f + ',' + ' '.join(map(str,o))+'\n' for f,o in zip(raw_filenames,trn_bbox)]

In [ ]:

open(f'{PATH}trn_bb_labels', 'w').writelines(trn_bb_labels)

In [ ]:

fnames,csv_labels,_,_ = parse_csv_labels(f'{PATH}trn_bb_labels', skip_header=False)

In [ ]:

def bb_corners(bb):
    bb = np.array(bb, dtype=np.float32)
    row1 = bb[3]
    col1 = bb[2]
    row2 = row1 + bb[0]
    col2 = col1 + bb[1]
    return [row1, col1, row2, col2]

In [ ]:

f = 'img_02642.jpg'
bb = csv_labels[f]
print(bb)
bb_corners(bb)

In [ ]:

new_labels = [f + "," + " ".join(map(str, bb_corners(csv_labels[f]))) + "\n" for f in raw_filenames]

In [ ]:

open(f'{PATH}trn_bb_corners_labels', 'w').writelines(new_labels)

Looking at bounding boxes¶

In [ ]:

# reading bb file
bbox = {}
bb_data = pd.read_csv(f'{PATH}trn_bb_labels', header=None)

In [ ]:

fnames,csv_labels,_,_ = parse_csv_labels(f'{PATH}trn_bb_labels', skip_header=False)
fnames,corner_labels,_,_ = parse_csv_labels(f'{PATH}trn_bb_corners_labels', skip_header=False)

In [ ]:

corner_labels["img_06297.jpg"]

Out[ ]:

['396.21', '404.67', '504.78', '623.22']

In [ ]:

csv_labels["img_06297.jpg"]

Out[ ]:

['108.57', '218.55', '404.67', '396.21']

In [ ]:

def create_rect(bb, color='red'):
    return plt.Rectangle((bb[2], bb[3]), bb[1], bb[0], color=color, fill=False, lw=3)

def show_bb(path, f='img_04908.jpg'):
    file_path = f'{path}images/{f}'
    bb = csv_labels[f]
    plots_from_files([file_path])
    plt.gca().add_patch(create_rect(bb))

In [ ]:

def create_corner_rect(bb, color='red'):
    bb = np.array(bb, dtype=np.float32)
    return plt.Rectangle((bb[1], bb[0]), bb[3]-bb[1], bb[2]-bb[0], color=color, fill=False, lw=3)

def show_corner_bb(path, f='img_04908.jpg'):
    file_path = f'{path}images/{f}'
    bb = corner_labels[f]
    plots_from_files([file_path])
    plt.gca().add_patch(create_corner_rect(bb))

In [ ]:

show_corner_bb(PATH, f = 'img_02642.jpg')

Model predicting bounding boxes¶

In [ ]:

sz=299
bs=64

label_csv=f'{PATH}trn_bb_corners_labels'
n = len(list(open(label_csv)))-1
val_idxs = get_cv_idxs(n)

In [ ]:

tfms = tfms_from_model(resnet34, sz, crop_type=CropType.NO, tfm_y=TfmType.COORD)
data = ImageClassifierData.from_csv(PATH, 'images', label_csv, tfms=tfms, val_idxs=val_idxs,
                                    continuous=True, skip_header=False)

In [ ]:

trn_ds = data.trn_dl.dataset

In [ ]:

x, y = trn_ds[0]

In [ ]:

print(x.shape, y)

(3, 299, 299) [ 56  62 163 150]

In [ ]:

learn = ConvLearner.pretrained(resnet34, data, precompute=True, opt_fn=optim.Adam, ps=0.5)

  0%|          | 0/48 [00:00<?, ?it/s]

---------------------------------------------------------------------------
KeyboardInterrupt                         Traceback (most recent call last)
<ipython-input-8-2599e6e01994> in <module>()
----> 1 learn = ConvLearner.pretrained(resnet34, data, precompute=True, opt_fn=optim.Adam, ps=0.5)

~/github/fastai/courses/dl1/fastai/conv_learner.py in pretrained(self, f, data, ps, xtra_fc, xtra_cut, **kwargs)
     87     def pretrained(self, f, data, ps=None, xtra_fc=None, xtra_cut=0, **kwargs):
     88         models = ConvnetBuilder(f, data.c, data.is_multi, data.is_reg, ps=ps, xtra_fc=xtra_fc, xtra_cut=xtra_cut)
---> 89         return self(data, models, **kwargs)
     90 
     91     @property

~/github/fastai/courses/dl1/fastai/conv_learner.py in __init__(self, data, models, precompute, **kwargs)
     80         elif self.metrics is None:
     81             self.metrics = [accuracy_multi] if self.data.is_multi else [accuracy]
---> 82         self.save_fc1()
     83         self.freeze()
     84         self.precompute=precompute

~/github/fastai/courses/dl1/fastai/conv_learner.py in save_fc1(self)
    121         if len(self.activations[0])==0:
    122             m=self.models.top_model
--> 123             predict_to_bcolz(m, self.data.fix_dl, act)
    124             predict_to_bcolz(m, self.data.val_dl, val_act)
    125             if self.data.test_dl: predict_to_bcolz(m, self.data.test_dl, test_act)

~/github/fastai/courses/dl1/fastai/model.py in predict_to_bcolz(m, gen, arr, workers)
      9     lock=threading.Lock()
     10     m.eval()
---> 11     for x,*_ in tqdm(gen):
     12         y = to_np(m(VV(x)).data)
     13         with lock:

~/anaconda/lib/python3.6/site-packages/tqdm/_tqdm.py in __iter__(self)
    870 """, fp_write=getattr(self.fp, 'write', sys.stderr.write))
    871 
--> 872             for obj in iterable:
    873                 yield obj
    874                 # Update and print the progressbar.

~/github/fastai/courses/dl1/fastai/dataset.py in __next__(self)
    218         if self.i>=len(self.dl): raise StopIteration
    219         self.i+=1
--> 220         return next(self.it)
    221 
    222     @property

~/anaconda/lib/python3.6/site-packages/torch/utils/data/dataloader.py in __next__(self)
    193         while True:
    194             assert (not self.shutdown and self.batches_outstanding > 0)
--> 195             idx, batch = self.data_queue.get()
    196             self.batches_outstanding -= 1
    197             if idx != self.rcvd_idx:

~/anaconda/lib/python3.6/queue.py in get(self, block, timeout)
    162             elif timeout is None:
    163                 while not self._qsize():
--> 164                     self.not_empty.wait()
    165             elif timeout < 0:
    166                 raise ValueError("'timeout' must be a non-negative number")

~/anaconda/lib/python3.6/threading.py in wait(self, timeout)
    293         try:    # restore state no matter what (e.g., KeyboardInterrupt)
    294             if timeout is None:
--> 295                 waiter.acquire()
    296                 gotit = True
    297             else:

KeyboardInterrupt:

In [ ]:

lrf=learn.lr_find()
learn.sched.plot()




  0%|          | 0/48 [00:00<?, ?it/s]



                                                         







[    0.       234.0974  4450.0396]

In [ ]:

learn.fit(0.01, 5, cycle_len=1, cycle_mult=2)



  0%|          | 0/48 [00:00<?, ?it/s]




                                                         







[   0.      238.681   364.7433]




  0%|          | 0/48 [00:00<?, ?it/s]



                                                         







[   1.      263.2261  186.7536]




  0%|          | 0/48 [00:00<?, ?it/s]



                                                         







[   2.      210.889   168.2116]




  0%|          | 0/48 [00:00<?, ?it/s]



                                                         







[   3.      176.0693  144.2513]




  0%|          | 0/48 [00:00<?, ?it/s]



                                                         







[   4.      151.4565  129.4032]




  0%|          | 0/48 [00:00<?, ?it/s]



                                                         







[   5.      138.8294  123.4013]



  0%|          | 0/48 [00:00<?, ?it/s]




                                                         







[   6.      131.2182  122.3169]



  0%|          | 0/48 [00:00<?, ?it/s]




                                                         







[   7.      125.9923  114.4574]



  0%|          | 0/48 [00:00<?, ?it/s]




                                                         







[   8.      121.7743  109.9652]



  0%|          | 0/48 [00:00<?, ?it/s]




                                                         







[   9.      117.3897  104.9769]



  0%|          | 0/48 [00:00<?, ?it/s]




                                                         







[  10.      112.0477  100.5813]



  0%|          | 0/48 [00:00<?, ?it/s]




                                                         







[  11.      109.0689   98.3695]




  0%|          | 0/48 [00:00<?, ?it/s]



                                                         





[  12.      106.7465   98.7911]
[  13.      105.5374   97.5796]                          
[  14.      105.4487   96.7256]                          
[  15.      104.3599   98.1003]                          
[  16.      106.321    95.0515]                          
[  17.      104.1545   94.0836]                          
[  18.      102.1962   92.3605]                          
[  19.      100.5235   91.6409]                          
[ 20.      99.355   89.3562]                              
[ 21.      98.3139  88.7288]                              
[ 22.      97.8664  87.1307]                              
[ 23.      95.7317  85.4799]                              
[ 24.      93.4525  86.0595]                              
[ 25.      93.5365  83.7847]                              
[ 26.      93.2249  84.8397]                              
[ 27.      93.2065  85.1662]                              
[ 28.      91.5837  83.8048]                              
[ 29.      90.2605  83.596 ]                              
[ 30.      92.2839  82.4151]

In [ ]:

lrs=np.array([1e-4,1e-3,1e-2])
learn.precompute=False

learn.unfreeze()
lrf=learn.lr_find(lrs/1e3)
learn.sched.plot()

[  0.0000e+00   1.3241e+02   3.5670e+13]

In [ ]:

lrs=np.array([1e-5,1e-4,1e-3])
learn.fit(lrs, 5, cycle_len=1, cycle_mult=2)

[  0.      71.1009  94.999 ]                             
[  1.      88.0754  82.1335]                             
[  2.      87.4359  77.1687]                              
[  3.      86.5974  80.8263]                              
[  4.      83.9792  71.4977]                              
[  5.      79.1773  69.3897]                              
[  6.      77.2359  68.047 ]                              
[  7.      77.6487  77.4456]                              
[  8.      78.9041  73.7645]                              
[  9.      75.8422  68.0503]                              
[ 10.      73.4623  64.6083]                              
[ 11.      72.2943  63.5834]                              
[ 12.      70.4287  62.1409]                              
[ 13.      66.5874  59.9783]                              
[ 14.      66.0414  60.785 ]                              
[ 15.      66.3314  65.7976]                              
[ 16.      69.3549  65.7114]                              
[ 17.      67.7767  64.7094]                              
[ 18.      67.158   61.4241]                              
[ 19.      65.6192  58.7482]                              
[ 20.      62.9917  60.5301]                              
[ 21.      64.713   61.0712]                              
[ 22.      61.309   56.2067]                              
[ 23.      58.6235  55.2833]                              
[ 24.      57.5971  55.6252]                              
[ 25.      58.6757  55.0441]                              
[ 26.      57.1917  52.6833]                              
[ 27.      56.4466  53.1598]                              
[ 28.      56.      51.4715]                              
[ 29.      55.0628  52.217 ]                              
[ 30.      55.1955  51.3541]

Looking into size of images¶

In [ ]:

f="img_06297.jpg"
PIL.Image.open(PATH+"images/" + f).size

In [ ]:

sizes = [PIL.Image.open(PATH+f).size for f in data.trn_ds.fnames]
raw_val_sizes = [PIL.Image.open(PATH+f).size for f in data.val_ds.fnames]