# Written by Francois Fleuret <francois@fleuret.org>
-import math, sys, argparse, time, tqdm, os
+import math, sys, argparse, time, tqdm, os, datetime
import torch, torchvision
from torch import nn
parser.add_argument("--grid_size", type=int, default=6)
+parser.add_argument("--grid_fraction_play", type=float, default=0)
+
##############################
# picoclvr options
"nb_test_samples": 10000,
},
"memory": {
- "model": "4M",
+ "model": "37M",
"batch_size": 100,
- "nb_train_samples": 5000,
+ "nb_train_samples": 25000,
"nb_test_samples": 1000,
},
"mixing": {
nb_test_samples=args.nb_test_samples,
batch_size=args.batch_size,
size=args.grid_size,
+ fraction_play=args.grid_fraction_play,
logger=log_string,
device=device,
)
deterministic_synthesis=args.deterministic_synthesis,
)
+time_pred_result = None
+
for n_epoch in range(nb_epochs_finished, nb_epochs):
learning_rate = learning_rate_schedule[n_epoch]
deterministic_synthesis=args.deterministic_synthesis,
)
+ time_current_result = datetime.datetime.now()
+ if time_pred_result is not None:
+ log_string(
+ f"next_result {time_current_result + (time_current_result - time_pred_result)}"
+ )
+ time_pred_result = time_current_result
+
checkpoint = {
"nb_epochs_finished": n_epoch + 1,
"model_state": model.state_dict(),