diff --git a/main.py b/main.py
index 2939f46a1b7b8c82c9c366e1f0115613c75a8367..4309779e648637e77e56235e23bcd8fac9382b49 100644
--- a/main.py
+++ b/main.py
@@ -1,6 +1,7 @@
import os
import torch
import torch.nn as nn
+import argparse
import numpy as np
import data
@@ -9,21 +10,32 @@ import musicgenerator as mg
device = "cuda" if torch.cuda.is_available() else "cpu"
if __name__ == '__main__':
+ parser = argparse.ArgumentParser()
+ parser.add_argument('--lr', type=float, default=20, help='initial learning rate')
+ parser.add_argument('--epochs', type=int, default=100, help='upper epoch limit')
+ parser.add_argument('--batch_size', type=int, default=16, help='batch size')
+ parser.add_argument('--sequence_length', type=int, default=32, help='sequence length')
+ parser.add_argument('--dimension_model', type=int, default=128, help='size of word embeddings')
+ parser.add_argument('--nhead', type=int, default=4, help='the number of heads in the encoder/decoder of the transformer model')
+ parser.add_argument('--dropout', type=float, default=0.2, help='dropout applied to layers (0 = no dropout)')
+ parser.add_argument('--nhid', type=int, default=200, help='number of hidden units per layer')
+ parser.add_argument('--nlayers', type=int, default=2, help='number of layers')
+ args = parser.parse_args()
###############################################################################
# Paramètres
###############################################################################
data_path = os.path.normpath('./training_data/classical_music/')
- learning_rate = 20
- batch_size = 16
+ learning_rate = args.lr
+ batch_size = args.batch_size
split_train_test_valid = (0.8, 0.1, 0.1)
model_path = None # os.path.join(data_path, 'model.pt')
- sequence_length = 32
- dim_model = 128
- num_head = 8
- num_layers = 2
- num_hid = 200
- dropout = 0.2
- epochs = 100
+ sequence_length = args.sequence_length
+ dim_model = args.dimension_model
+ num_head = args.nhead
+ num_layers = args.nlayers
+ num_hid = args.nhid
+ dropout = args.dropout
+ epochs = args.epochs
nb_log_epoch = 5
nb_words = 100
temperature = 1.0
diff --git a/parameter_finder.py b/parameter_finder.py
new file mode 100644
index 0000000000000000000000000000000000000000..d98fc516f0785d6d9b1471b52a49324c29cee7a3
--- /dev/null
+++ b/parameter_finder.py
@@ -0,0 +1,61 @@
+from matplotlib import pyplot as plt
+import subprocess
+import itertools
+import time
+import re
+
+lr = ['0.1', '1', '100']
+epochs = ['20', '50', '100']
+batch_size = ['4', '16', '64']
+sequence_length = ['8', '32', '128']
+dimension_model = ['64', '256', '512']
+nhead = ['2', '4', '8']
+dropout = ['0.0', '0.3', '0.6']
+nhid = ['100', '200', '500']
+nlayers = ['2', '6', '10']
+
+if __name__ == '__main__':
+ best_ppl = None
+ best_config = ''
+ nb_tests = 0
+ for lr_i, epoch_i, batch_size_i, sequence_length_i, dimension_model_i, nhead_i, dropout_i, nhid_i, nlayers_i in itertools.product(
+ lr, epochs, batch_size, sequence_length, dimension_model, nhead, dropout, nhid, nlayers):
+ res_name = f"lr{str.replace(lr_i, '.', '')}_epoch{epoch_i}_batch{batch_size_i}_seq{sequence_length_i}_dim{dimension_model_i}_nhead{nhead_i}_drop{str.replace(dropout_i, '.', '')}_nhid{nhid_i}_nlay{nlayers_i}"
+ nb_tests += 1
+ print("Start config :", res_name)
+ start_time = time.time()
+ res = subprocess.run(['python', 'main.py',
+ '--lr', lr_i,
+ '--epochs', epoch_i,
+ '--batch_size', batch_size_i,
+ '--sequence_length', sequence_length_i,
+ '--dimension_model', dimension_model_i,
+ '--nhead', nhead_i,
+ '--dropout', dropout_i,
+ '--nhid', nhid_i,
+ '--nlayers', nlayers_i
+ ],
+ stdout=subprocess.PIPE,
+ stderr=subprocess.PIPE
+ )
+ end_time = time.time()
+ res_stdout = res.stdout.decode()
+ res_stderr = res.stderr.decode()
+ match = re.search(r'test ppl\s+([0-9.]+)', res_stdout)
+ if match:
+
+ test_ppl = float(match.group(1))
+
+ if best_ppl is None or test_ppl < best_ppl:
+ best_ppl = test_ppl
+ best_config = res_name
+
+ print(res_name, test_ppl, "time :", end_time - start_time)
+ else:
+ print("ERROR:", res_stdout, res_stderr)
+ if nb_tests >= 15:
+ break
+
+ print("nb tests :", nb_tests)
+ print("best config :", best_config)
+ print("best ppl :", best_ppl)
diff --git a/test.py b/test.py
index 5f6a8e0917e5693e6ccb74518b175e62bf8efb1a..36647c2dfd39a86a395d8616864612976521ce44 100644
--- a/test.py
+++ b/test.py
@@ -1,4 +1,5 @@
import argparse
+import math
import random
if __name__ == '__main__':
@@ -9,10 +10,10 @@ if __name__ == '__main__':
parser.add_argument('--batch_size', type=int, default=16, help='batch size')
parser.add_argument('--sequence_length', type=int, default=32, help='sequence length')
parser.add_argument('--dimension_model', type=int, default=128, help='size of word embeddings')
-
parser.add_argument('--nhead', type=int, default=4, help='the number of heads in the encoder/decoder of the transformer model')
parser.add_argument('--dropout', type=float, default=0.2, help='dropout applied to layers (0 = no dropout)')
parser.add_argument('--nhid', type=int, default=200, help='number of hidden units per layer')
parser.add_argument('--nlayers', type=int, default=2, help='number of layers')
-
+ test_loss = random.uniform(0.0, 15.0)
+ print('| End of training | test loss {:5.2f} | test ppl {:8.2f}'.format(test_loss, math.exp(test_loss)))
diff --git a/training_data/classical_music/output.mid b/training_data/classical_music/output.mid
index d809b8e8481dec1ff304d4fe97dea41c32ae0c64..dce21e73e0425bb7ebbb5322f415d2394683aaf7 100644
Binary files a/training_data/classical_music/output.mid and b/training_data/classical_music/output.mid differ