keraTorch 0.0.2

A keras like wrapper for pytorch

KeraTorch

Implementing Keras clone with pytorch backend.

Install

pip install keratorch

How to use

from keraTorch.model import Sequential
from keraTorch.layers import *
from keraTorch.losses import *

The data:

x_train.shape, y_train.shape, x_valid.shape, y_valid.shape

((50000, 784), (50000,), (10000, 784), (10000,))

Model definition:

model = Sequential()
model.add(Dense(100, x_train.shape[1], activation='relu'))
model.add(Dense(50, activation='relu'))
model.add(Dense(10))
model.add(Activation('softmax'))

Doesn't actually compile anything but to look like keras we specify the loss as:

model.compile(ce4softmax)

Burrow for Fastai's learning rate finder to find best learning rate:

bs = 256
model.lr_find(x_train, y_train, bs=bs)

<div>
    <style>
        /* Turns off some styling */
        progress {
            /* gets rid of default border in Firefox and Opera. */
            border: none;
            /* Needs to be in here for Safari polyfill so background images work as expected. */
            background-size: auto;
        }
        .progress-bar-interrupted, .progress-bar-interrupted::-webkit-progress-bar {
            background: #F44336;
        }
    </style>
  <progress value='25' class='' max='34' style='width:300px; height:20px; vertical-align: middle;'></progress>
  73.53% [25/34 00:00<00:00]
</div>

epoch	train_loss	valid_loss	time
0	2.301613	#na#	00:00
1	2.301608	#na#	00:00
2	2.301600	#na#	00:00
3	2.301588	#na#	00:00
4	2.301569	#na#	00:00
5	2.301540	#na#	00:00
6	2.301493	#na#	00:00
7	2.301417	#na#	00:00
8	2.301294	#na#	00:00
9	2.301092	#na#	00:00
10	2.300761	#na#	00:00
11	2.300216	#na#	00:00
12	2.299308	#na#	00:00
13	2.297771	#na#	00:00
14	2.295127	#na#	00:00
15	2.290489	#na#	00:00
16	2.281993	#na#	00:00
17	2.265558	#na#	00:00
18	2.230882	#na#	00:00
19	2.157919	#na#	00:00
20	2.041476	#na#	00:00
21	1.920061	#na#	00:00
22	1.823919	#na#	00:00
23	1.768780	#na#	00:00
24	1.723490	#na#	00:00
25	1.641395	#na#	00:00
26	1.776127	#na#	00:00
27	2.319300	#na#	00:00
28	3.339199	#na#	00:00
29	4.307324	#na#	00:00
30	5.229871	#na#	00:00
31	6.128597	#na#	00:00

<div>
    <style>
        /* Turns off some styling */
        progress {
            /* gets rid of default border in Firefox and Opera. */
            border: none;
            /* Needs to be in here for Safari polyfill so background images work as expected. */
            background-size: auto;
        }
        .progress-bar-interrupted, .progress-bar-interrupted::-webkit-progress-bar {
            background: #F44336;
        }
    </style>
  <progress value='1' class='' max='3' style='width:300px; height:20px; vertical-align: middle;'></progress>
  33.33% [1/3 00:00<00:00 6.4181]
</div>



LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.
Min numerical gradient: 9.12E-03
Min loss divided by 10: 1.74E-02

We have the same .fit and .predict functions:

model.fit(x_train, y_train, bs, epochs=5, lr=1e-2)

epoch	train_loss	valid_loss	time
0	2.294942	2.212783	00:01
1	2.163815	1.622664	00:01
2	1.856739	1.071931	00:01
3	1.548884	0.879702	00:01
4	1.325464	0.846312	00:01

preds = model.predict(x_valid)
accuracy = (preds.argmax(axis=-1) == y_valid).mean()
print(f'Predicted accuracy is {accuracy:.2f}')

Predicted accuracy is 0.77