-
Notifications
You must be signed in to change notification settings - Fork 397
/
plot_hessian_eigen.py
245 lines (207 loc) · 11.6 KB
/
plot_hessian_eigen.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
"""
Calculate the hessian matrix of the projected surface and their eigen values.
"""
import argparse
import copy
import numpy as np
import h5py
import torch
import time
import socket
import os
import sys
import torchvision
import torch.nn as nn
import mpi4pytorch
import dataloader
import net_plotter
import plot_2D
import plot_1D
import model_loader
import scheduler
import projection as proj
import hess_vec_prod
from plot_surface import name_surface_file, setup_surface_file
def crunch_hessian_eigs(surf_file, net, w, s, d, dataloader, comm, rank, args):
"""
Calculate eigen values of the hessian matrix of a given model in parallel
using mpi reduce. This is the synchronized version.
"""
f = h5py.File(surf_file, 'r+' if rank == 0 else 'r')
min_eig, max_eig = [], []
xcoordinates = f['xcoordinates'][:]
ycoordinates = f['ycoordinates'][:] if 'ycoordinates' in f.keys() else None
if 'min_eig' not in f.keys():
shape = xcoordinates.shape if ycoordinates is None else (len(xcoordinates),len(ycoordinates))
max_eig = -np.ones(shape=shape)
min_eig = np.ones(shape=shape)
if rank == 0:
f['min_eig'] = min_eig
f['max_eig'] = max_eig
else:
min_eig = f['min_eig'][:]
max_eig = f['max_eig'][:]
# Generate a list of all indices that need to be filled in.
# The coordinates of each unfilled index are stored in 'coords'.
inds, coords, inds_nums = scheduler.get_job_indices(max_eig, xcoordinates, ycoordinates, comm)
print('Computing %d values for rank %d'% (len(inds), rank))
criterion = nn.CrossEntropyLoss() # set the loss function criteria
# Loop over all un-calculated coords
start_time = time.time()
total_sync = 0.0
for count, ind in enumerate(inds):
# Get the coordinates of the points being calculated
coord = coords[count]
# Load the weights corresponding to those coordinates into the net
if args.dir_type == 'weights':
net_plotter.set_weights(net.module if args.ngpu > 1 else net, w, d, coord)
elif args.dir_type == 'states':
net_plotter.set_states(net.module if args.ngpu > 1 else net, s, d, coord)
# Compute the eign values of the hessian matrix
compute_start = time.time()
maxeig, mineig, iter_count = hess_vec_prod.min_max_hessian_eigs(net, dataloader, \
criterion, rank=rank, use_cuda=args.cuda, verbose=True)
compute_time = time.time() - compute_start
# Record the result in the local array
max_eig.ravel()[ind] = maxeig
min_eig.ravel()[ind] = mineig
# Send updated plot data to the master node
sync_start_time = time.time()
max_eig = mpi4pytorch.reduce_max(comm, max_eig)
min_eig = mpi4pytorch.reduce_min(comm, min_eig)
sync_time = time.time() - sync_start_time
total_sync += sync_time
# Only the master node writes to the file - this avoids write conflicts
if rank == 0:
f['max_eig'][:] = max_eig
f['min_eig'][:] = min_eig
print("rank: %d %d/%d (%0.2f%%) %d\t %s \tmaxeig:%8.5f \tmineig:%8.5f \titer: %d \ttime:%.2f \tsync:%.2f" % ( \
rank, count + 1, len(inds), 100.0 * (count + 1)/len(inds), ind, str(coord), \
maxeig, mineig, iter_count, compute_time, sync_time))
# This is only needed to make MPI run smoothly. If this process has less work
# than the rank0 process, then we need to keep calling allreduce so the rank0 process doesn't block
for i in range(max(inds_nums) - len(inds)):
max_eig = mpi4pytorch.reduce_max(comm, max_eig)
min_eig = mpi4pytorch.reduce_min(comm, min_eig)
total_time = time.time() - start_time
print('Rank %d done! Total time: %f Sync: %f '%(rank, total_time, total_sync))
f.close()
###############################################################
#### MAIN
###############################################################
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='plotting loss surface')
parser.add_argument('--mpi', '-m', action='store_true', help='use mpi')
parser.add_argument('--cuda', '-c', action='store_true', help='use cuda')
parser.add_argument('--threads', default=2, type=int, help='number of threads')
parser.add_argument('--ngpu', type=int, default=1, help='number of GPUs to use for each rank, useful for data parallel evaluation')
parser.add_argument('--batch_size', default=128, type=int, help='minibatch size')
# data parameters
parser.add_argument('--dataset', default='cifar10', help='cifar10 | imagenet')
parser.add_argument('--datapath', default='cifar10/data', metavar='DIR', help='path to the dataset')
parser.add_argument('--raw_data', action='store_true', default=False, help='no data preprocessing')
parser.add_argument('--data_split', default=1, type=int, help='the number of splits for the dataloader')
parser.add_argument('--split_idx', default=0, type=int, help='the index of data splits for the dataloader')
parser.add_argument('--trainloader', default='', help='path to the dataloader with random labels')
parser.add_argument('--testloader', default='', help='path to the testloader with random labels')
# model parameters
parser.add_argument('--model', default='resnet56', help='model name')
parser.add_argument('--model_folder', default='', help='the common folder that contains model_file and model_file2')
parser.add_argument('--model_file', default='', help='path to the trained model file')
parser.add_argument('--model_file2', default='', help='use (model_file2 - model_file) as the xdirection')
parser.add_argument('--model_file3', default='', help='use (model_file3 - model_file) as the ydirection')
parser.add_argument('--loss_name', '-l', default='crossentropy', help='loss functions: crossentropy | mse')
# direction parameters
parser.add_argument('--dir_file', default='', help='specify the name of direction file, or the path to an eisting direction file')
parser.add_argument('--dir_type', default='weights', help='direction type: weights | states (including BN\'s running_mean/var)')
parser.add_argument('--x', default='-1:1:51', help='A string with format xmin:x_max:xnum')
parser.add_argument('--y', default=None, help='A string with format ymin:ymax:ynum')
parser.add_argument('--xnorm', default='', help='direction normalization: filter | layer | weight')
parser.add_argument('--ynorm', default='', help='direction normalization: filter | layer | weight')
parser.add_argument('--xignore', default='', help='ignore bias and BN parameters: biasbn')
parser.add_argument('--yignore', default='', help='ignore bias and BN parameters: biasbn')
parser.add_argument('--idx', default=0, type=int, help='the index for the repeatness experiment')
parser.add_argument('--surf_file', default='', help='customize the name of surface file, could be an existing file.')
# plot parameters
parser.add_argument('--show', action='store_true', default=False, help='show plotted figures')
parser.add_argument('--plot', action='store_true', default=False, help='plot figures after computation')
args = parser.parse_args()
torch.manual_seed(123)
#--------------------------------------------------------------------------
# Environment setup
#--------------------------------------------------------------------------
if args.mpi:
comm = mpi4pytorch.setup_MPI()
rank, nproc = comm.Get_rank(), comm.Get_size()
else:
comm, rank, nproc = None, 0, 1
# in case of multiple GPUs per node, set the GPU to use for each rank
if args.cuda:
if not torch.cuda.is_available():
raise Exception('User selected cuda option, but cuda is not available on this machine')
gpu_count = torch.cuda.device_count()
torch.cuda.set_device(rank % gpu_count)
print('Rank %d use GPU %d of %d GPUs on %s' %
(rank, torch.cuda.current_device(), gpu_count, socket.gethostname()))
#--------------------------------------------------------------------------
# Check plotting resolution
#--------------------------------------------------------------------------
try:
args.xmin, args.xmax, args.xnum = [float(a) for a in args.x.split(':')]
args.ymin, args.ymax, args.ynum = (None, None, None)
if args.y:
args.ymin, args.ymax, args.ynum = [float(a) for a in args.y.split(':')]
assert args.ymin and args.ymax and args.ynum, \
'You specified some arguments for the y axis, but not all'
except:
raise Exception('Improper format for x- or y-coordinates. Try something like -1:1:51')
#--------------------------------------------------------------------------
# Load models and extract parameters
#--------------------------------------------------------------------------
net = model_loader.load(args.dataset, args.model, args.model_file)
w = net_plotter.get_weights(net) # initial parameters
s = copy.deepcopy(net.state_dict()) # deepcopy since state_dict are references
if args.ngpu > 1:
# data parallel with multiple GPUs on a single node
net = nn.DataParallel(net, device_ids=range(torch.cuda.device_count()))
#--------------------------------------------------------------------------
# Setup the direction file and the surface file
#--------------------------------------------------------------------------
dir_file = net_plotter.name_direction_file(args) # name the direction file
if rank == 0:
net_plotter.setup_direction(args, dir_file, net)
surf_file = name_surface_file(args, dir_file)
if rank == 0:
setup_surface_file(args, surf_file, dir_file)
# wait until master has setup the direction file and surface file
mpi4pytorch.barrier(comm)
# load directions
d = net_plotter.load_directions(dir_file)
# calculate the consine similarity of the two directions
if len(d) == 2 and rank == 0:
similarity = proj.cal_angle(proj.nplist_to_tensor(d[0]), proj.nplist_to_tensor(d[1]))
print('cosine similarity between x-axis and y-axis: %f' % similarity)
#--------------------------------------------------------------------------
# Setup dataloader
#--------------------------------------------------------------------------
# download CIFAR10 if it does not exit
if rank == 0 and args.dataset == 'cifar10':
torchvision.datasets.CIFAR10(root=args.dataset + '/data', train=True, download=True)
mpi4pytorch.barrier(comm)
trainloader, testloader = dataloader.load_dataset(args.dataset, args.datapath,
args.batch_size, args.threads, args.raw_data,
args.data_split, args.split_idx,
args.trainloader, args.testloader)
#--------------------------------------------------------------------------
# Start the computation
#--------------------------------------------------------------------------
crunch_hessian_eigs(surf_file, net, w, s, d, trainloader, comm, rank, args)
print ("Rank " + str(rank) + ' is done!')
#--------------------------------------------------------------------------
# Plot figures
#--------------------------------------------------------------------------
if args.plot and rank == 0:
if args.y:
plot_2D.plot_2d_eig_ratio(surf_file, 'min_eig', 'max_eig', args.show)
else:
plot_1D.plot_1d_eig_ratio(surf_file, args.xmin, args.xmax, 'min_eig', 'max_eig')