Skip to content

Commit a9ed181

Browse files
author
Ilia Karmanov
committed
higher-lvel APIs
1 parent 5139830 commit a9ed181

File tree

2 files changed

+684
-0
lines changed

2 files changed

+684
-0
lines changed

Diff for: ‎notebooks/CNTK_CNN_highAPI.ipynb

+301
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,301 @@
1+
{
2+
"cells": [
3+
{
4+
"cell_type": "markdown",
5+
"metadata": {},
6+
"source": [
7+
"# High-level CNTK Example"
8+
]
9+
},
10+
{
11+
"cell_type": "code",
12+
"execution_count": 1,
13+
"metadata": {},
14+
"outputs": [],
15+
"source": [
16+
"import numpy as np\n",
17+
"import os\n",
18+
"import sys\n",
19+
"import cntk\n",
20+
"from cntk.layers import Convolution2D, MaxPooling, Dense, Dropout\n",
21+
"from common.params import *\n",
22+
"from common.utils import *"
23+
]
24+
},
25+
{
26+
"cell_type": "code",
27+
"execution_count": 2,
28+
"metadata": {},
29+
"outputs": [],
30+
"source": [
31+
"# Force one-gpu\n",
32+
"os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"0\""
33+
]
34+
},
35+
{
36+
"cell_type": "code",
37+
"execution_count": 3,
38+
"metadata": {},
39+
"outputs": [
40+
{
41+
"name": "stdout",
42+
"output_type": "stream",
43+
"text": [
44+
"OS: linux\n",
45+
"Python: 3.5.2 |Anaconda custom (64-bit)| (default, Jul 2 2016, 17:53:06) \n",
46+
"[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)]\n",
47+
"Numpy: 1.14.1\n",
48+
"CNTK: 2.4\n",
49+
"GPU: ['Tesla P100-PCIE-16GB', 'Tesla P100-PCIE-16GB']\n",
50+
"CUDA Version 8.0.61\n",
51+
"CuDNN Version 6.0.21\n"
52+
]
53+
}
54+
],
55+
"source": [
56+
"print(\"OS: \", sys.platform)\n",
57+
"print(\"Python: \", sys.version)\n",
58+
"print(\"Numpy: \", np.__version__)\n",
59+
"print(\"CNTK: \", cntk.__version__)\n",
60+
"print(\"GPU: \", get_gpu_name())\n",
61+
"print(get_cuda_version())\n",
62+
"print(\"CuDNN Version \", get_cudnn_version())"
63+
]
64+
},
65+
{
66+
"cell_type": "code",
67+
"execution_count": 4,
68+
"metadata": {},
69+
"outputs": [],
70+
"source": [
71+
"def create_symbol(n_classes=N_CLASSES):\n",
72+
" # Weight initialiser from uniform distribution\n",
73+
" # Activation (unless states) is None\n",
74+
" with cntk.layers.default_options(init = cntk.glorot_uniform(), activation = cntk.relu):\n",
75+
" x = Convolution2D(filter_shape=(3, 3), num_filters=50, pad=True)(features)\n",
76+
" x = Convolution2D(filter_shape=(3, 3), num_filters=50, pad=True)(x)\n",
77+
" x = MaxPooling((2, 2), strides=(2, 2), pad=False)(x)\n",
78+
" x = Dropout(0.25)(x)\n",
79+
"\n",
80+
" x = Convolution2D(filter_shape=(3, 3), num_filters=100, pad=True)(x)\n",
81+
" x = Convolution2D(filter_shape=(3, 3), num_filters=100, pad=True)(x)\n",
82+
" x = MaxPooling((2, 2), strides=(2, 2), pad=False)(x)\n",
83+
" x = Dropout(0.25)(x) \n",
84+
" \n",
85+
" x = Dense(512)(x)\n",
86+
" x = Dropout(0.5)(x)\n",
87+
" x = Dense(n_classes, activation=None)(x)\n",
88+
" return x"
89+
]
90+
},
91+
{
92+
"cell_type": "code",
93+
"execution_count": 5,
94+
"metadata": {},
95+
"outputs": [],
96+
"source": [
97+
"def init_model(m, labels, lr=LR, momentum=MOMENTUM):\n",
98+
" # Loss (dense labels); check if support for sparse labels\n",
99+
" loss = cntk.cross_entropy_with_softmax(m, labels)\n",
100+
" # Momentum SGD\n",
101+
" # https://github.com/Microsoft/CNTK/blob/master/Manual/Manual_How_to_use_learners.ipynb\n",
102+
" # unit_gain=False: momentum_direction = momentum*old_momentum_direction + gradient\n",
103+
" # if unit_gain=True then ...(1-momentum)*gradient\n",
104+
" learner = cntk.momentum_sgd(m.parameters, \n",
105+
" lr=cntk.learning_rate_schedule(lr, cntk.UnitType.minibatch) , \n",
106+
" momentum=cntk.momentum_schedule(momentum),\n",
107+
" unit_gain=False)\n",
108+
" return loss, learner"
109+
]
110+
},
111+
{
112+
"cell_type": "code",
113+
"execution_count": 6,
114+
"metadata": {},
115+
"outputs": [
116+
{
117+
"name": "stdout",
118+
"output_type": "stream",
119+
"text": [
120+
"Preparing train set...\n",
121+
"Preparing test set...\n",
122+
"(50000, 3, 32, 32) (10000, 3, 32, 32) (50000, 10) (10000, 10)\n",
123+
"float32 float32 float32 float32\n",
124+
"CPU times: user 738 ms, sys: 575 ms, total: 1.31 s\n",
125+
"Wall time: 1.31 s\n"
126+
]
127+
}
128+
],
129+
"source": [
130+
"%%time\n",
131+
"# Data into format for library\n",
132+
"x_train, x_test, y_train, y_test = cifar_for_library(channel_first=True, one_hot=True)\n",
133+
"# CNTK format\n",
134+
"y_train = y_train.astype(np.float32)\n",
135+
"y_test = y_test.astype(np.float32)\n",
136+
"print(x_train.shape, x_test.shape, y_train.shape, y_test.shape)\n",
137+
"print(x_train.dtype, x_test.dtype, y_train.dtype, y_test.dtype)"
138+
]
139+
},
140+
{
141+
"cell_type": "code",
142+
"execution_count": 7,
143+
"metadata": {},
144+
"outputs": [
145+
{
146+
"name": "stdout",
147+
"output_type": "stream",
148+
"text": [
149+
"CPU times: user 16.7 ms, sys: 40.4 ms, total: 57.1 ms\n",
150+
"Wall time: 67.4 ms\n"
151+
]
152+
}
153+
],
154+
"source": [
155+
"%%time\n",
156+
"# Placeholders\n",
157+
"features = cntk.input_variable((3, 32, 32), np.float32)\n",
158+
"labels = cntk.input_variable(N_CLASSES, np.float32)\n",
159+
"# Load symbol\n",
160+
"sym = create_symbol()"
161+
]
162+
},
163+
{
164+
"cell_type": "code",
165+
"execution_count": 8,
166+
"metadata": {},
167+
"outputs": [
168+
{
169+
"name": "stdout",
170+
"output_type": "stream",
171+
"text": [
172+
"CPU times: user 122 ms, sys: 116 ms, total: 238 ms\n",
173+
"Wall time: 239 ms\n"
174+
]
175+
}
176+
],
177+
"source": [
178+
"%%time\n",
179+
"loss, learner = init_model(sym, labels)"
180+
]
181+
},
182+
{
183+
"cell_type": "code",
184+
"execution_count": 9,
185+
"metadata": {},
186+
"outputs": [
187+
{
188+
"name": "stdout",
189+
"output_type": "stream",
190+
"text": [
191+
"CPU times: user 37.9 s, sys: 10.8 s, total: 48.7 s\n",
192+
"Wall time: 48.8 s\n"
193+
]
194+
},
195+
{
196+
"data": {
197+
"text/plain": [
198+
"{'epoch_summaries': [{'loss': 1.8144259375, 'metric': 0.0, 'samples': 50000},\n",
199+
" {'loss': 1.36322234375, 'metric': 0.0, 'samples': 50000},\n",
200+
" {'loss': 1.122504140625, 'metric': 0.0, 'samples': 50000},\n",
201+
" {'loss': 0.974794296875, 'metric': 0.0, 'samples': 50000},\n",
202+
" {'loss': 0.8672890625, 'metric': 0.0, 'samples': 50000},\n",
203+
" {'loss': 0.7853078125, 'metric': 0.0, 'samples': 50000},\n",
204+
" {'loss': 0.716815546875, 'metric': 0.0, 'samples': 50000},\n",
205+
" {'loss': 0.65541078125, 'metric': 0.0, 'samples': 50000},\n",
206+
" {'loss': 0.606273671875, 'metric': 0.0, 'samples': 50000},\n",
207+
" {'loss': 0.560514921875, 'metric': 0.0, 'samples': 50000}],\n",
208+
" 'updates': [{'loss': 1.8144589081005922, 'metric': 0.0, 'samples': 49984},\n",
209+
" {'loss': 1.363123699583867, 'metric': 0.0, 'samples': 49984},\n",
210+
" {'loss': 1.1224501996889005, 'metric': 0.0, 'samples': 49984},\n",
211+
" {'loss': 0.9746546238546335, 'metric': 0.0, 'samples': 49984},\n",
212+
" {'loss': 0.8671638205475752, 'metric': 0.0, 'samples': 49984},\n",
213+
" {'loss': 0.7853081736155569, 'metric': 0.0, 'samples': 49984},\n",
214+
" {'loss': 0.7168769787582027, 'metric': 0.0, 'samples': 49984},\n",
215+
" {'loss': 0.6554993300981314, 'metric': 0.0, 'samples': 49984},\n",
216+
" {'loss': 0.6063771656930218, 'metric': 0.0, 'samples': 49984},\n",
217+
" {'loss': 0.5606013064805738, 'metric': 0.0, 'samples': 49984}]}"
218+
]
219+
},
220+
"execution_count": 9,
221+
"metadata": {},
222+
"output_type": "execute_result"
223+
}
224+
],
225+
"source": [
226+
"%%time\n",
227+
"# Main training loop: 49s\n",
228+
"loss.train((x_train, y_train), \n",
229+
" minibatch_size=BATCHSIZE, \n",
230+
" max_epochs=EPOCHS,\n",
231+
" parameter_learners=[learner])"
232+
]
233+
},
234+
{
235+
"cell_type": "code",
236+
"execution_count": 10,
237+
"metadata": {},
238+
"outputs": [
239+
{
240+
"name": "stdout",
241+
"output_type": "stream",
242+
"text": [
243+
"CPU times: user 284 ms, sys: 95.9 ms, total: 380 ms\n",
244+
"Wall time: 409 ms\n"
245+
]
246+
}
247+
],
248+
"source": [
249+
"%%time\n",
250+
"# Main evaluation loop: 409ms\n",
251+
"n_samples = (y_test.shape[0]//BATCHSIZE)*BATCHSIZE\n",
252+
"y_guess = np.zeros(n_samples, dtype=np.int)\n",
253+
"y_truth = np.argmax(y_test[:n_samples], axis=-1)\n",
254+
"c = 0\n",
255+
"for data, label in yield_mb(x_test, y_test, BATCHSIZE):\n",
256+
" predicted_label_probs = sym.eval({features : data})\n",
257+
" y_guess[c*BATCHSIZE:(c+1)*BATCHSIZE] = np.argmax(predicted_label_probs, axis=-1)\n",
258+
" c += 1"
259+
]
260+
},
261+
{
262+
"cell_type": "code",
263+
"execution_count": 11,
264+
"metadata": {},
265+
"outputs": [
266+
{
267+
"name": "stdout",
268+
"output_type": "stream",
269+
"text": [
270+
"Accuracy: 0.7591145833333334\n"
271+
]
272+
}
273+
],
274+
"source": [
275+
"print(\"Accuracy: \", 1.*sum(y_guess == y_truth)/len(y_guess))"
276+
]
277+
}
278+
],
279+
"metadata": {
280+
"anaconda-cloud": {},
281+
"kernelspec": {
282+
"display_name": "Python 3",
283+
"language": "python",
284+
"name": "python3"
285+
},
286+
"language_info": {
287+
"codemirror_mode": {
288+
"name": "ipython",
289+
"version": 3
290+
},
291+
"file_extension": ".py",
292+
"mimetype": "text/x-python",
293+
"name": "python",
294+
"nbconvert_exporter": "python",
295+
"pygments_lexer": "ipython3",
296+
"version": "3.5.2"
297+
}
298+
},
299+
"nbformat": 4,
300+
"nbformat_minor": 2
301+
}

0 commit comments

Comments
 (0)