-
Notifications
You must be signed in to change notification settings - Fork 2
/
Tutorial.py
381 lines (271 loc) · 12.9 KB
/
Tutorial.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
# coding: utf-8
# In[1]:
import numpy as np
import pprint as pp
from sklearn.model_selection import train_test_split
import pickle
import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable
import torch.utils.data as data_utils
import matplotlib.pylab as plt
import sys, os
sys.path.append(os.path.join(os.path.dirname("__file__"), '..'))
from pytorch_net.net import MLP, ConvNet, load_model_dict, train, get_Layer
from pytorch_net.util import Early_Stopping, get_param_name_list, get_variable_name_list, standardize_symbolic_expression
# ### Apart from section 0, each section is independent on its own
# ## 0. Preparing dataset:
# In[2]:
# Preparing some toy dataset:
X = np.random.randn(1000,1)
y = X ** 2
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.2)
X_train = Variable(torch.FloatTensor(X_train))
y_train = Variable(torch.FloatTensor(y_train))
X_test = Variable(torch.FloatTensor(X_test))
y_test = Variable(torch.FloatTensor(y_test))
# ## 1. Constuct a simple MLP:
# In[3]:
# Constuct the network:
input_size = 1
struct_param = [
[2, "Simple_Layer", {}], # (number of neurons in each layer, layer_type, layer settings)
[400, "Simple_Layer", {"activation": "relu"}],
[1, "Simple_Layer", {"activation": "linear"}],
]
settings = {"activation": "relu"} # Default activation if the activation is not specified in "struct_param" in each layer.
# If the activation is specified, it will overwrite this default settings.
net = MLP(input_size = input_size,
struct_param = struct_param,
settings = settings,
)
# In[ ]:
# Get the prediction of the Net:
net(X_train)
# In[ ]:
# Get intermediate activation of the net:
net.inspect_operation(X_train, # Input
operation_between = (0,2), # Operation_between selects the subgraph.
# Here (0, 2) means that the inputs feeds into layer_0 (first layer)
# and before the layer_2 (third layer).
)
# In[ ]:
# Save model:
pickle.dump(net.model_dict, open("net.p", "wb")) # net.model_dict constains all the information (structural and parameters) of the network
# Load model:
net_loaded = load_model_dict(pickle.load(open("net.p", "rb")))
# Check the loaded net and the original net is identical:
net_loaded(X_train) - net(X_train)
# ## 2. Using symbolic layers, and simplification:
# ### 2.1 Constructing MLP consisting of symbolic layers:
# In[ ]:
# Construct the network:
model_dict = {
"type": "MLP",
"input_size": 4,
"struct_param": [[2, "Symbolic_Layer", {"symbolic_expression": "[3 * x0 ** 2 + p0 * x1 * x2 + p1 * x3, 5 * x0 ** 2 + p2 * x1 + p3 * x3 * x2]"}],
[1, "Symbolic_Layer", {"symbolic_expression": "[3 * x0 ** 2 + p2 * x1]"}],
],
# Here the optional "weights" sets up the initial values for the parameters. If not set, will initialize with N(0, 1):
'weights': [{'p0': -1.3,
'p1': 1.0,
'p2': 2.3,
'p3': -0.4},
{'p2': -1.5},
]
}
net = load_model_dict(model_dict)
pp.pprint(net.model_dict)
print("\nOutput:")
net(torch.rand(100, 4))
# ### 2.2 Simplification of an MLP from Simple_Layer to Symbolic_Layer:
# In[6]:
input_size = 1
struct_param = [
[2, "Simple_Layer", {"activation": "relu"}], # (number of neurons in each layer, layer_type, layer settings)
[10, "Simple_Layer", {"activation": "linear"}],
[1, "Simple_Layer", {"activation": "relu"}],
]
net = MLP(input_size = input_size,
struct_param = struct_param,
settings = {},
)
# In[ ]:
net.simplify(X=X_train,
y=y_train,
mode=['to_symbolic'],
# The mode is a list of consecutive simplification methods, choosing from:
# 'collapse_layers': collapse multiple Simple_Layer with linear activation into a single Simple_Layer;
# 'local': greedily try reducing the input dimension by removing input dimension from the beginning
# 'snap': greedily snap each float parameter into an integer or rational number. Set argument 'snap_mode' == 'integer' or 'rational';
# 'pair_snap': greedily trying if the ratio of a pair of parameters is an integer or rational number (by setting snap_mode)
# 'activation_snap': snap the activation;
# 'to_symbolic': transform the Simple_Layer into Symbolic_layer;
# 'symbolic_simplification': collapse multiple layers of Symbolic_Layer into a single Symbolic_Layer;
# 'ramping-L1': increasing L1 regularization for the parameters and train. When some parameter is below a threshold, snap it to 0.
)
# ## 3. SuperNet:
# ### 3.1 Use SuperNet Layer in your own module:
# In[7]:
layer_dict = {
"layer_type": "SuperNet_Layer",
"input_size": 1,
"output_size": 2,
"settings": {
"W_available": ["dense", "Toeplitz"], # Weight type. Choose subset of "dense", "Toeplitz", "arithmetic-series-in", "arithmetic-series-out", "arithmetic-series-2D-in", "arithmetic-series-2D-out"
"b_available": ["dense", "None", "arithmetic-series", "constant"], # Bias type. Choose subsets of "None", "constant", "arithmetic-series", "arithmetic-series-2D"
"A_available": ["linear", "relu"], # Activation. Choose subset of "linear", "relu", "leakyRelu", "softplus", "sigmoid", "tanh", "selu", "elu", "softmax"
}
}
SuperNet_Layer = get_Layer(**layer_dict)
# In[8]:
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.layer_0 = SuperNet_Layer
self.layer_1 = nn.Linear(2,4)
def forward(self, X):
return self.layer_1(self.layer_0(X))
# In[ ]:
net = Net()
net(X_train)
# ### 3.2 Construct an MLP containing SuperNet layer:
# In[ ]:
input_size = 1
struct_param = [
[6, "Simple_Layer", {}], # (number of neurons in each layer, layer_type, layer settings)
[4, "SuperNet_Layer", {"activation": "relu", # Choose from "linear", "relu", "leakyRelu", "softplus", "sigmoid", "tanh", "selu", "elu", "softmax"
"W_available": ["dense", "Toeplitz"],
"b_available": ["dense", "None", "arithmetic-series", "constant"],
"A_available": ["linear", "relu"],
}],
[1, "Simple_Layer", {"activation": "linear"}],
]
settings = {"activation": "relu"} # Default activation if the activation is not specified in "struct_param" in each layer.
# If the activation is specified, it will overwrite this default settings.
net = MLP(input_size = input_size,
struct_param = struct_param,
settings = settings,
)
net(X_train)
# ## 4. Training using explicit commands:
# In[34]:
# training settings:
batch_size = 128
epochs = 500
# Prepare training set batches:
dataset_train = data_utils.TensorDataset(X_train.data, y_train.data) # The data_loader must use the torch Tensor, not Variable. So I use X_train.data to get the Tensor.
train_loader = data_utils.DataLoader(dataset_train, batch_size = batch_size, shuffle = True)
# Set up optimizer:
optimizer = optim.Adam(net.parameters(), lr = 1e-3)
# Get loss function. Choose from "mse" or "huber", etc.
criterion = nn.MSELoss()
# Set up early stopping. If the validation loss does not go down after "patience" number of epochs, early stop.
early_stopping = Early_Stopping(patience = 10)
# In[ ]:
to_stop = False
for epoch in range(epochs):
for batch_id, (X_batch, y_batch) in enumerate(train_loader):
# Every learning step must contain the following 5 steps:
optimizer.zero_grad() # Zero out the gradient buffer
pred = net(Variable(X_batch)) # Obtain network's prediction
loss_train = criterion(pred, Variable(y_batch)) # Calculate the loss
loss_train.backward() # Perform backward step on the loss to calculate the gradient for each parameter
optimizer.step() # Use the optimizer to perform one step of parameter update
# Validation at the end of each epoch:
loss_test = criterion(net(X_test), y_test)
to_stop = early_stopping.monitor(loss_test.item())
print("epoch {0} \tbatch {1} \tloss_train: {2:.6f}\tloss_test: {3:.6f}".format(epoch, batch_id, loss_train.item(), loss_test.item()))
if to_stop:
print("Early stopping at epoch {0}".format(epoch))
break
# In[6]:
# Save model:
pickle.dump(net.model_dict, open("net.p", "wb"))
# Load model:
net_loaded = load_model_dict(pickle.load(open("net.p", "rb")))
# Check the loaded net and the original net is identical:
net_loaded(X_train) - net(X_train)
# ## 5. Advanced example: training MNIST using given train() function:
# In[ ]:
from torchvision import datasets
import torch.utils.data as data_utils
from pytorch_net.util import train_test_split, normalize_tensor, to_Variable
is_cuda = torch.cuda.is_available()
struct_param_conv = [
[64, "Conv2d", {"kernel_size": 3, "padding": 1}],
[64, "BatchNorm2d", {"activation": "relu"}],
[None, "MaxPool2d", {"kernel_size": 2}],
[64, "Conv2d", {"kernel_size": 3, "padding": 1}],
[64, "BatchNorm2d", {"activation": "relu"}],
[None, "MaxPool2d", {"kernel_size": 2}],
[64, "Conv2d", {"kernel_size": 3, "padding": 1}],
[64, "BatchNorm2d", {"activation": "relu"}],
[None, "MaxPool2d", {"kernel_size": 2}],
[64, "Conv2d", {"kernel_size": 3, "padding": 1}],
[64, "BatchNorm2d", {"activation": "relu"}],
[None, "MaxPool2d", {"kernel_size": 2}],
[10, "Simple_Layer", {"layer_input_size": 64}],
]
model = ConvNet(input_channels = 1,
struct_param = struct_param_conv,
is_cuda = is_cuda,
)
# In[ ]:
dataset_name = "MNIST"
batch_size = 128
dataset_raw = getattr(datasets, dataset_name)('datasets/{0}'.format(dataset_name), download = True)
X, y = to_Variable(dataset_raw.train_data.unsqueeze(1).float(), dataset_raw.train_labels, is_cuda = is_cuda)
X = normalize_tensor(X, new_range = (0, 1))
(X_train, y_train), (X_test, y_test) = train_test_split(X, y, test_size = 0.2) # Split into training and testing
dataset_train = data_utils.TensorDataset(X_train, y_train)
train_loader = data_utils.DataLoader(dataset_train, batch_size = batch_size, shuffle = True) # initialize the dataLoader
# In[ ]:
lr = 1e-3
reg_dict = {"weight": 1e-6, "bias": 1e-6}
loss_original, loss_value, data_record = train(model,
train_loader = train_loader,
validation_data = (X_test, y_test),
criterion = nn.CrossEntropyLoss(),
lr = lr,
reg_dict = reg_dict,
epochs = 1000,
isplot = True,
patience = 40,
scheduler_patience = 40,
inspect_items = ["accuracy"],
record_keys = ["accuracy"],
inspect_interval = 1,
inspect_items_interval = 1,
inspect_loss_precision = 4,
)
# ## 6. An example callback code:
# In[ ]:
if isplot:
import matplotlib.pylab as plt
fig = plt.figure()
if dim == 3:
ax = fig.add_subplot(111, projection='3d')
else:
fig.add_axes([0, 0, 1, 1])
ax = fig.axes[0]
else:
ax = None
def visualize(model, X, y, iteration, loss, ax, dim):
if isplot:
plt.cla()
y_pred = model.transform(X)
y_pred, y = to_np_array(y_pred, y)
if dim == 3:
ax.scatter(y[:,0], y[:,1], y[:,2], color='red', label='ref', s = 1)
ax.scatter(y_pred[:,0], y_pred[:,1], y_pred[:,2], color='blue', label='data', s = 1)
ax.text2D(0.87, 0.92, 'Iteration: {:d}\nError: {:06.4f}'.format(iteration, loss), horizontalalignment='center', verticalalignment='center', transform=ax.transAxes, fontsize='x-large')
else:
ax.scatter(y[:,0], y[:,1], color='red', label='ref')
ax.scatter(y_pred[:,0], y_pred[:,1], color='blue', label='data')
plt.text(0.87, 0.92, 'Iteration: {:d}\nError: {:06.4f}'.format(iteration, loss), horizontalalignment='center', verticalalignment='center', transform=ax.transAxes, fontsize='x-large')
ax.legend(loc='upper left', fontsize='x-large')
plt.draw()
plt.pause(pause)
callback = partial(visualize, ax = ax, dim = dim)