-
Notifications
You must be signed in to change notification settings - Fork 0
/
train001.json
376 lines (255 loc) · 10.8 KB
/
train001.json
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
#%%
import time
import jax
## Use jax cpu
# jax.config.update("jax_platform_name", "cpu")
import jax.numpy as jnp
import numpy as np
import equinox as eqx
import optax
import matplotlib.pyplot as plt
from functools import partial
import datetime
# from flax.metrics import tensorboard
from nodepint.utils import get_new_keys, sbplot, seconds_to_hours
from nodepint.training import train_project_neural_ode, test_neural_ode
# from nodepint.data import load_jax_dataset, get_dataset_features, preprocess_mnist
from nodepint.data import load_mnist_dataset_torch
from nodepint.integrators import dopri_integrator, euler_integrator, rk4_integrator, dopri_integrator_diff
from nodepint.pint import newton_root_finder, direct_root_finder, fixed_point_finder, direct_root_finder_aug, parareal
from nodepint.sampling import random_sampling, identity_sampling, neural_sampling
import cProfile
import os
# os.environ['XLA_FLAGS'] = '--xla_force_host_platform_device_count=4' ## Trick to virtualise CPU for pmap
print("Available devices:", jax.devices())
# os.environ['XLA_FLAGS'] = '--xla_gpu_force_compilation_parallelism=1' ## For things to work on JADE (single-threaded compilation)
# nb_devices = jax.local_device_count()
SEED = 27
#%% [markdown]
# ## Define neural net
#%%
# class MLP(eqx.Module):
# """
# A simple neural net that learn MNIST
# """
# layers: list
# # prediction_layer: eqx.nn.Linear
# def __init__(self, key=None):
# key = get_new_keys(key)
# self.layers = [eqx.nn.Linear(100, 100, key=key)]
# for i in range(1):
# self.layers = self.layers + [jax.nn.relu, eqx.nn.Linear(100, 100, key=key)]
# # self.prediction_layer = eqx.nn.Linear(100, 10, key=key)
# def __call__(self, x):
# for layer in self.layers:
# x = layer(x)
# return x
class Encoder(eqx.Module):
"""
A convolutional encoder for MNIST
"""
layers: list
def __init__(self, key=None):
keys = get_new_keys(key, num=3)
self.layers = [eqx.nn.Conv2d(1, 64, (3, 3), stride=1, key=keys[0]), jax.nn.relu, eqx.nn.GroupNorm(64, 64),
eqx.nn.Conv2d(64, 64, (4, 4), stride=2, padding=1, key=keys[1]), jax.nn.relu, eqx.nn.GroupNorm(64, 64),
eqx.nn.Conv2d(64, 64, (4, 4), stride=2, padding=1, key=keys[2]) ]
def __call__(self, x):
for layer in self.layers:
x = layer(x)
return x
class Processor(eqx.Module):
"""
A convlutional processor to be passed to the neural ODE
"""
layers: list
def __init__(self, key=None):
keys = get_new_keys(key, num=2)
self.layers = [eqx.nn.Conv2d(64+1, 64, (3, 3), stride=1, padding=1, key=keys[0]), jax.nn.tanh,
eqx.nn.Conv2d(64, 64, (3, 3), stride=1, padding=1, key=keys[1]), jax.nn.tanh]
def __call__(self, x, t):
y = jnp.concatenate([jnp.broadcast_to(t, (1,)+x.shape[1:]), x], axis=0)
for layer in self.layers:
y = layer(y)
return y
class Decoder(eqx.Module):
"""
A decoder to classify MNIST
"""
layers: list
def __init__(self, key=None):
key = get_new_keys(key, 1)
self.layers = [eqx.nn.GroupNorm(64, 64), jax.nn.relu,
eqx.nn.AvgPool2d((6, 6)), lambda x:jnp.reshape(x, (64,)),
eqx.nn.Linear(64, 10, key=key)]
def __call__(self, x):
for layer in self.layers:
x = layer(x)
return x
#%% [markdown]
# ## Load the dataset
#%%
ds = load_mnist_dataset_torch(root="./data/mnist", train=True)
# ds = make_dataloader_torch(ds, subset_size="all", seed=SEED, norm_factor=255.)
# print("Feature names:", get_dataset_features(ds))
print("Number of training examples:", len(ds))
## Visualise a datapoint
np.random.seed(time.time_ns()%(2**32))
point_id = np.random.randint(0, len(ds))
pixels, label = ds[point_id]
plt.title(f"Label is {label:1d}")
plt.imshow(pixels.squeeze(), cmap='gray')
plt.show()
# import torch
# print("Torch devices out there", torch.cuda.device_count())
#%% [markdown]
# ## Define training parameters
#%%
## Optax crossentropy loss
optim_scheme = optax.adam
# times = tuple(np.linspace(0, 1, 101).flatten())
times = (0., 1., 101) ## t0, tf, nb_times (this is for solving the ODE if an adaptative time stepper is not used. Not for eval)
integrator_args = (1e-1, 1e-1, jnp.inf, 10, 2, "checkpointed") ## rtol, atol, max_dt, max_steps, kind, max_steps_rev (these are typically by adatative time steppers)
fixed_point_args = (1., 1e-6, 10) ## learning_rate, tol, max_iter
# loss = optax.softmax_cross_entropy
loss = optax.softmax_cross_entropy_with_integer_labels
# def cross_entropy_fn(y_pred, y): ## TODO: should be vmapped by design
# y_pred = jnp.argmax(jax.nn.softmax(y_pred, axis=-1), axis=-1)
# y = jnp.argmax(y, axis=-1)
# return jnp.mean(y_pred == y, axis=-1)
## Base neural ODE model
# neuralnet = MLP(key=SEED)
# neuralnet = eqx.nn.MLP(in_size=100, out_size=100, width_size=250, depth=3, activation=jax.nn.relu, key=get_key(None))
keys = get_new_keys(SEED, num=3)
neural_nets = (Encoder(key=keys[0]), Processor(key=keys[1]), Decoder(key=keys[2]))
## PinT scheme with only mandatory arguments
nb_epochs = 100
batch_size = 120*1 ## Divisible by the dataset size to avoid recompilation !
total_steps = nb_epochs*(len(ds)//batch_size)
scheduler = optax.piecewise_constant_schedule(init_value=1e-4, boundaries_and_scales={int(total_steps*0.5):0.25, int(total_steps*0.75):0.25})
key = get_new_keys(SEED)
train_params = {"neural_nets":neural_nets,
"data":ds,
# "pint_scheme":fixed_point_finder,
"pint_scheme":parareal,
"samp_scheme":neural_sampling,
# "samp_scheme":identity_sampling,
# "integrator":rk4_integrator,
# "integrator":euler_integrator,
# "integrator":dopri_integrator,
"integrator":dopri_integrator_diff,
"integrator_args":integrator_args,
"loss_fn":loss,
"optim_scheme":optim_scheme,
"nb_processors":800,
"scheduler":scheduler,
"times":times,
"fixed_point_args":fixed_point_args,
"nb_epochs":nb_epochs,
"batch_size":batch_size,
"repeat_projection":1,
"nb_vectors":5,
"force_serial":True,
"key":key}
#%% [markdown]
# ## Train the model
#%%
# with jax.profiler.trace("./runs", create_perfetto_link=False):
# profiler = cProfile.Profile()
# profiler.enable()
start_time = time.time()
cpu_start_time = time.process_time()
trained_networks, shooting_fn, loss_hts, errors_hts, nb_iters_hts = train_project_neural_ode(**train_params)
clock_time = time.process_time() - cpu_start_time
wall_time = time.time() - start_time
# print("\nNumber of iterations till PinT eventual convergence:\n", np.asarray(nb_iters_hts))
# print("Errors during PinT iterations:\n", np.asarray(errors_hts))
time_in_hmsecs = seconds_to_hours(wall_time)
print("\nTotal training time: %d hours %d mins %d secs" %time_in_hmsecs)
# profiler.disable()
# # profiler.print_stats(sort='cumulative')
# profile_output_filename = "runs/cprofile/profile_report.txt"
# with open(profile_output_filename, "w") as f:
# profiler.dump_stats(profile_output_filename)
#%%
# eqx.tree_serialise_leaves("data/encode_process_decode.eqx", neural_nets)
#%%
# trained_networks = eqx.tree_deserialise_leaves("data/encode_process_decode.eqx", neural_nets)
# shooting_fn, loss_hts, errors_hts, nb_iters_hts = [None]*4
#%% [markdown]
# ## Analyse loss history
#%%
## Plot the loss histories per iterations
labels = [str(i) for i in range(len(loss_hts))]
epochs = range(len(loss_hts[0]))
sbplot(epochs, jnp.stack(loss_hts, axis=-1), label=labels, x_label="epochs", y_scale="log", title="Loss histories");
## Loss histories acros all iterations
total_loss = np.concatenate(loss_hts, axis=0)
total_epochs = 1 + np.arange(len(total_loss))
ax = sbplot(total_epochs, total_loss, x_label="epochs", y_scale="log", title="Total loss history");
# #%% [markdown]
# # ## Quick profiling
# #%%
# import cProfile
# def main():
# train_parallel_neural_ode(**train_params)
# cProfile.run('main()', sort='cumtime')
#%% [markdown]
# ## Compute metrics on a test dataset
#%%
## Load the test dataset
test_ds = load_mnist_dataset_torch(root="./data/mnist", train=False)
print("\nNumber of testing examples", len(test_ds))
def accuracy_fn(y_pred, y):
y_pred = jnp.argmax(jax.nn.softmax(y_pred, axis=-1), axis=-1)
# y = jnp.argmax(y, axis=-1)
return jnp.sum(y_pred == y, axis=-1)*100
test_params = {"neural_nets": trained_networks,
"data":test_ds,
"pint_scheme":fixed_point_finder, ## If None then the fixed_point_ad_rule is used
# "pint_scheme":direct_scheme,
# "integrator":rk4_integrator,
"integrator":dopri_integrator_diff,
"integrator_args":integrator_args,
"fixed_point_args":fixed_point_args,
"acc_fn":accuracy_fn,
"shooting_fn":shooting_fn,
"nb_processors":16,
"times":times,
"batch_size":120}
start_time = time.time()
avg_acc = test_neural_ode(**test_params)
test_wall_time = time.time() - start_time
time_in_hms= seconds_to_hours(test_wall_time)
print(f"\nAverage accuracy: {avg_acc:.2f} %")
print("Test time: %d hours %d mins %d secs" %time_in_hms)
# #%% [markdown]
# # ## Write stuff to tensorboard
# #%%
# run_name = str(datetime.datetime.now().strftime("%H:%M %d-%m-%Y"))[:19]
# writer = tensorboard.SummaryWriter("runs/"+run_name)
# hps = {}
# hps["bach_size"] = train_params["batch_size"]
# hps["scheduler"] = train_params["scheduler"]
# hps["nb_epochs"] = train_params["nb_epochs"]
# hps["nb_processors"] = train_params["nb_processors"]
# hps["repeat_projection"] = train_params["repeat_projection"]
# hps["nb_vectors"] = train_params["nb_vectors"]
# hps["times"] = (train_params["times"][0], train_params["times"][-1], len(train_params["times"]))
# hps["optim_scheme"] = train_params["optim_scheme"].__name__
# hps["pint_scheme"] = str(train_params["pint_scheme"])[45:-63]
# hps["key"] = SEED
# hps["integrator"] = train_params["integrator"].__name__
# hps["loss_fn"] = train_params["loss_fn"].__name__
# hps["data"] = str(get_dataset_features(train_params["data"]))
# hps["dynamicnet_size"] = sum(x.size for x in jax.tree_util.tree_leaves(eqx.partition(dynamicnet, eqx.is_array)[0]))
# hps["wall_time"] = wall_time
# hps["clock_time"] = clock_time
# hps["test_acc"] = avg_acc
# hps["test_wall_time"] = test_wall_time
# writer.hparams(hps)
# for ep in range(len(total_epochs)):
# writer.scalar('train_loss', total_loss[ep], ep+1)
# writer.flush()
# writer.close()
# %%