In [1]:
from __future__ import division, print_function
%matplotlib inline
#path = "data/state/"
path = "data/state/sample/"
from importlib import reload # Python 3
import utils; reload(utils)
from utils import *
from IPython.display import FileLink
In [2]:
batch_size=64
In [3]:
batches = get_batches(path+'train', batch_size=batch_size)
val_batches = get_batches(path+'valid', batch_size=batch_size*2, shuffle=False)
steps_per_epoch = int(np.ceil(batches.samples/batch_size))
validation_steps = int(np.ceil(val_batches.samples/(batch_size*2)))
In [4]:
(val_classes, trn_classes, val_labels, trn_labels,
val_filenames, filenames, test_filenames) = get_classes(path)
Rather than using batches, we could just import all the data into an array to save some processing time. (In most examples I'm using the batches, however - just because that's how I happened to start out.)
In [5]:
trn = get_data(path+'train')
val = get_data(path+'valid')
In [6]:
save_array(path+'results/val.dat', val)
save_array(path+'results/trn.dat', trn)
In [7]:
val = load_array(path+'results/val.dat')
trn = load_array(path+'results/trn.dat')
We should find that everything that worked on the sample (see statefarm-sample.ipynb), works on the full dataset too. Only better! Because now we have more data. So let's see how they go - the models in this section are exact copies of the sample notebook models.
In [8]:
def conv1(batches):
model = Sequential([
BatchNormalization(axis=1, input_shape=(3,224,224)),
Conv2D(32,(3,3), activation='relu'),
BatchNormalization(axis=1),
MaxPooling2D((3,3)),
Conv2D(64,(3,3), activation='relu'),
BatchNormalization(axis=1),
MaxPooling2D((3,3)),
Flatten(),
Dense(200, activation='relu'),
BatchNormalization(),
Dense(10, activation='softmax')
])
model.compile(Adam(lr=1e-4), loss='categorical_crossentropy', metrics=['accuracy'])
model.fit_generator(batches, steps_per_epoch, epochs=2, validation_data=val_batches,
validation_steps=validation_steps)
model.optimizer.lr = 0.001
model.fit_generator(batches, steps_per_epoch, epochs=4, validation_data=val_batches,
validation_steps=validation_steps)
return model
In [9]:
model = conv1(batches)
Interestingly, with no regularization or augmentation we're getting some reasonable results from our simple convolutional model. So with augmentation, we hopefully will see some very good results.
In [10]:
gen_t = image.ImageDataGenerator(rotation_range=15, height_shift_range=0.05,
shear_range=0.1, channel_shift_range=20, width_shift_range=0.1)
batches = get_batches(path+'train', gen_t, batch_size=batch_size)
In [11]:
model = conv1(batches)
In [12]:
model.optimizer.lr = 0.0001
model.fit_generator(batches, steps_per_epoch, epochs=15, validation_data=val_batches,
validation_steps=validation_steps)
Out[12]:
I'm shocked by how good these results are! We're regularly seeing 75-80% accuracy on the validation set, which puts us into the top third or better of the competition. With such a simple model and no dropout or semi-supervised learning, this really speaks to the power of this approach to data augmentation.
Unfortunately, the results are still very unstable - the validation accuracy jumps from epoch to epoch. Perhaps a deeper model with some dropout would help.
In [13]:
gen_t = image.ImageDataGenerator(rotation_range=15, height_shift_range=0.05,
shear_range=0.1, channel_shift_range=20, width_shift_range=0.1)
batches = get_batches(path+'train', gen_t, batch_size=batch_size)
In [14]:
model = Sequential([
BatchNormalization(axis=1, input_shape=(3,224,224)),
Conv2D(32,(3,3), activation='relu'),
BatchNormalization(axis=1),
MaxPooling2D(),
Conv2D(64,(3,3), activation='relu'),
BatchNormalization(axis=1),
MaxPooling2D(),
Conv2D(128,(3,3), activation='relu'),
BatchNormalization(axis=1),
MaxPooling2D(),
Flatten(),
Dense(200, activation='relu'),
BatchNormalization(),
Dropout(0.5),
Dense(200, activation='relu'),
BatchNormalization(),
Dropout(0.5),
Dense(10, activation='softmax')
])
In [15]:
model.compile(Adam(lr=10e-5), loss='categorical_crossentropy', metrics=['accuracy'])
In [16]:
model.fit_generator(batches, steps_per_epoch, epochs=2, validation_data=val_batches,
validation_steps=validation_steps)
Out[16]:
In [17]:
model.optimizer.lr=0.001
In [18]:
model.fit_generator(batches, steps_per_epoch, epochs=10, validation_data=val_batches,
validation_steps=validation_steps)
Out[18]:
In [19]:
model.optimizer.lr=0.00001
In [20]:
model.fit_generator(batches, steps_per_epoch, epochs=10, validation_data=val_batches,
validation_steps=validation_steps)
Out[20]:
This is looking quite a bit better - the accuracy is similar, but the stability is higher. There's still some way to go however...
Since we have so little data, and it is similar to imagenet images (full color photos), using pre-trained VGG weights is likely to be helpful - in fact it seems likely that we won't need to fine-tune the convolutional layer weights much, if at all. So we can pre-compute the output of the last convolutional layer, as we did in lesson 3 when we experimented with dropout. (However this means that we can't use full data augmentation, since we can't pre-compute something that changes every image.)
In [21]:
vgg = Vgg16()
model=vgg.model
last_conv_idx = [i for i,l in enumerate(model.layers) if type(l) is Convolution2D][-1]
conv_layers = model.layers[:last_conv_idx+1]
In [22]:
conv_model = Sequential(conv_layers)
In [23]:
(val_classes, trn_classes, val_labels, trn_labels,
val_filenames, filenames, test_filenames) = get_classes(path)
In [24]:
test_batches = get_batches(path+'test', batch_size=batch_size*2, shuffle=False)
In [25]:
conv_feat = conv_model.predict_generator(batches, int(np.ceil(batches.samples/batch_size)))
conv_val_feat = conv_model.predict_generator(val_batches, int(np.ceil(val_batches.samples/(batch_size*2))))
conv_test_feat = conv_model.predict_generator(test_batches, int(np.ceil(test_batches.samples/(batch_size*2))))
In [26]:
save_array(path+'results/conv_val_feat.dat', conv_val_feat)
save_array(path+'results/conv_test_feat.dat', conv_test_feat)
save_array(path+'results/conv_feat.dat', conv_feat)
In [27]:
conv_feat = load_array(path+'results/conv_feat.dat')
conv_val_feat = load_array(path+'results/conv_val_feat.dat')
conv_val_feat.shape
Out[27]:
Since we've pre-computed the output of the last convolutional layer, we need to create a network that takes that as input, and predicts our 10 classes. Let's try using a simplified version of VGG's dense layers.
In [28]:
def get_bn_layers(p):
return [
MaxPooling2D(input_shape=conv_layers[-1].output_shape[1:]),
Flatten(),
Dropout(p/2),
Dense(128, activation='relu'),
BatchNormalization(),
Dropout(p/2),
Dense(128, activation='relu'),
BatchNormalization(),
Dropout(p),
Dense(10, activation='softmax')
]
In [29]:
p=0.8
In [30]:
bn_model = Sequential(get_bn_layers(p))
bn_model.compile(Adam(lr=0.001), loss='categorical_crossentropy', metrics=['accuracy'])
In [31]:
bn_model.fit(conv_feat, trn_labels, batch_size=batch_size, epochs=1,
validation_data=(conv_val_feat, val_labels))
Out[31]:
In [32]:
bn_model.optimizer.lr=0.01
In [33]:
bn_model.fit(conv_feat, trn_labels, batch_size=batch_size, epochs=2,
validation_data=(conv_val_feat, val_labels))
Out[33]:
In [34]:
bn_model.save_weights(path+'models/conv8.h5')
Looking good! Let's try pre-computing 5 epochs worth of augmented data, so we can experiment with combining dropout and augmentation on the pre-trained model.
We'll use our usual data augmentation parameters:
In [35]:
gen_t = image.ImageDataGenerator(rotation_range=15, height_shift_range=0.05,
shear_range=0.1, channel_shift_range=20, width_shift_range=0.1)
da_batches = get_batches(path+'train', gen_t, batch_size=batch_size, shuffle=False)
We use those to create a dataset of convolutional features 5x bigger than the training set.
In [36]:
da_conv_feat = conv_model.predict_generator(da_batches, 5*int(np.ceil((da_batches.samples)/(batch_size))), workers=3)
In [37]:
save_array(path+'results/da_conv_feat2.dat', da_conv_feat)
In [38]:
da_conv_feat = load_array(path+'results/da_conv_feat2.dat')
Let's include the real training data as well in its non-augmented form.
In [39]:
da_conv_feat = np.concatenate([da_conv_feat, conv_feat])
Since we've now got a dataset 6x bigger than before, we'll need to copy our labels 6 times too.
In [40]:
da_trn_labels = np.concatenate([trn_labels]*6)
Based on some experiments the previous model works well, with bigger dense layers.
In [41]:
def get_bn_da_layers(p):
return [
MaxPooling2D(input_shape=conv_layers[-1].output_shape[1:]),
Flatten(),
Dropout(p),
Dense(256, activation='relu'),
BatchNormalization(),
Dropout(p),
Dense(256, activation='relu'),
BatchNormalization(),
Dropout(p),
Dense(10, activation='softmax')
]
In [42]:
p=0.8
In [43]:
bn_model = Sequential(get_bn_da_layers(p))
bn_model.compile(Adam(lr=0.001), loss='categorical_crossentropy', metrics=['accuracy'])
Now we can train the model as usual, with pre-computed augmented data.
In [44]:
bn_model.fit(da_conv_feat, da_trn_labels, batch_size=batch_size, epochs=1,
validation_data=(conv_val_feat, val_labels))
Out[44]:
In [45]:
bn_model.optimizer.lr=0.01
In [46]:
bn_model.fit(da_conv_feat, da_trn_labels, batch_size=batch_size, epochs=4,
validation_data=(conv_val_feat, val_labels))
Out[46]:
In [47]:
bn_model.optimizer.lr=0.0001
In [48]:
bn_model.fit(da_conv_feat, da_trn_labels, batch_size=batch_size, epochs=4,
validation_data=(conv_val_feat, val_labels))
Out[48]:
Looks good - let's save those weights.
In [49]:
bn_model.save_weights(path+'models/da_conv8_1.h5')
We're going to try using a combination of pseudo labeling and knowledge distillation to allow us to use unlabeled data (i.e. do semi-supervised learning). For our initial experiment we'll use the validation set as the unlabeled data, so that we can see that it is working without using the test set. At a later date we'll try using the test set.
To do this, we simply calculate the predictions of our model...
In [50]:
val_pseudo = bn_model.predict(conv_val_feat, batch_size=batch_size)
...concatenate them with our training labels...
In [51]:
comb_pseudo = np.concatenate([da_trn_labels, val_pseudo])
In [52]:
comb_feat = np.concatenate([da_conv_feat, conv_val_feat])
...and fine-tune our model using that data.
In [53]:
bn_model.load_weights(path+'models/da_conv8_1.h5')
In [54]:
bn_model.fit(comb_feat, comb_pseudo, batch_size=batch_size, epochs=1,
validation_data=(conv_val_feat, val_labels))
Out[54]:
In [55]:
bn_model.fit(comb_feat, comb_pseudo, batch_size=batch_size, epochs=4,
validation_data=(conv_val_feat, val_labels))
Out[55]:
In [56]:
bn_model.optimizer.lr=0.00001
In [57]:
bn_model.fit(comb_feat, comb_pseudo, batch_size=batch_size, epochs=4,
validation_data=(conv_val_feat, val_labels))
Out[57]:
That's a distinct improvement - even although the validation set isn't very big. This looks encouraging for when we try this on the test set.
In [58]:
bn_model.save_weights(path+'models/bn-ps8.h5')
We'll find a good clipping amount using the validation set, prior to submitting.
In [59]:
def do_clip(arr, mx): return np.clip(arr, (1-mx)/9, mx)
In [60]:
val_preds = bn_model.predict(conv_val_feat, batch_size=batch_size*2)
In [61]:
np.mean(keras.metrics.categorical_crossentropy(val_labels, do_clip(val_preds, 0.93)).eval())
Out[61]:
In [62]:
conv_test_feat = load_array(path+'results/conv_test_feat.dat')
In [63]:
preds = bn_model.predict(conv_test_feat, batch_size=batch_size*2)
In [64]:
subm = do_clip(preds,0.93)
In [65]:
subm_name = path+'results/subm.gz'
In [66]:
classes = sorted(batches.class_indices, key=batches.class_indices.get)
In [67]:
submission = pd.DataFrame(subm, columns=classes)
submission.insert(0, 'img', [a[4:] for a in test_filenames])
submission.head()
Out[67]:
In [68]:
submission.to_csv(subm_name, index=False, compression='gzip')
In [69]:
FileLink(subm_name)
Out[69]:
This gets 0.534 on the leaderboard.
You can safely ignore everything from here on, because they didn't really help.
In [70]:
#for l in get_bn_layers(p): conv_model.add(l) # this choice would give a weight shape error
for l in get_bn_da_layers(p): conv_model.add(l) # ... so probably this is the right one
In [71]:
for l1,l2 in zip(bn_model.layers, conv_model.layers[last_conv_idx+1:]):
l2.set_weights(l1.get_weights())
In [72]:
for l in conv_model.layers: l.trainable =False
In [73]:
for l in conv_model.layers[last_conv_idx+1:]: l.trainable =True
In [74]:
comb = np.concatenate([trn, val])
In [75]:
# not knowing what the experiment was about, added this to avoid a shape match error with comb using gen_t.flow
comb_pseudo = np.concatenate([trn_labels, val_pseudo])
In [76]:
gen_t = image.ImageDataGenerator(rotation_range=8, height_shift_range=0.04,
shear_range=0.03, channel_shift_range=10, width_shift_range=0.08)
In [77]:
batches = gen_t.flow(comb, comb_pseudo, batch_size=batch_size)
In [78]:
val_batches = get_batches(path+'valid', batch_size=batch_size*2, shuffle=False)
In [79]:
conv_model.compile(Adam(lr=0.00001), loss='categorical_crossentropy', metrics=['accuracy'])
In [80]:
conv_model.fit_generator(batches, steps_per_epoch, epochs=1, validation_data=val_batches,
validation_steps=validation_steps)
Out[80]:
In [81]:
conv_model.optimizer.lr = 0.0001
In [82]:
conv_model.fit_generator(batches, steps_per_epoch, epochs=3, validation_data=val_batches,
validation_steps=validation_steps)
Out[82]:
In [83]:
for l in conv_model.layers[16:]: l.trainable =True
In [84]:
#- added compile instruction in order to avoid Keras 2.1 warning message
conv_model.compile(Adam(), loss='categorical_crossentropy', metrics=['accuracy'])
In [85]:
conv_model.optimizer.lr = 0.00001
In [86]:
conv_model.fit_generator(batches, steps_per_epoch, epochs=8, validation_data=val_batches,
validation_steps=validation_steps)
Out[86]:
In [87]:
conv_model.save_weights(path+'models/conv8_ps.h5')
In [88]:
#conv_model.load_weights(path+'models/conv8_da.h5') # conv8_da.h5 was not saved in this notebook
In [89]:
val_pseudo = conv_model.predict(val, batch_size=batch_size*2)
In [90]:
save_array(path+'models/pseudo8_da.dat', val_pseudo)
In [91]:
drivers_ds = pd.read_csv(path+'driver_imgs_list.csv')
drivers_ds.head()
Out[91]:
In [92]:
img2driver = drivers_ds.set_index('img')['subject'].to_dict()
In [93]:
driver2imgs = {k: g["img"].tolist()
for k,g in drivers_ds[['subject', 'img']].groupby("subject")}
In [94]:
# It seems this function is not used in this notebook
def get_idx(driver_list):
return [i for i,f in enumerate(filenames) if img2driver[f[3:]] in driver_list]
In [95]:
# drivers = driver2imgs.keys() # Python 2
drivers = list(driver2imgs) # Python 3
In [96]:
rnd_drivers = np.random.permutation(drivers)
In [97]:
ds1 = rnd_drivers[:len(rnd_drivers)//2]
ds2 = rnd_drivers[len(rnd_drivers)//2:]
In [ ]:
# The following cells seem to require some preparation code not included in this notebook
models=[fit_conv([d]) for d in drivers]
models=[m for m in models if m is not None]
In [ ]:
all_preds = np.stack([m.predict(conv_test_feat, batch_size=128) for m in models])
avg_preds = all_preds.mean(axis=0)
avg_preds = avg_preds/np.expand_dims(avg_preds.sum(axis=1), 1)
In [ ]:
keras.metrics.categorical_crossentropy(val_labels, np.clip(avg_val_preds,0.01,0.99)).eval()
In [ ]:
keras.metrics.categorical_accuracy(val_labels, np.clip(avg_val_preds,0.01,0.99)).eval()
In [ ]: