| 1 | import os |
| 2 | import sys |
| 3 | |
| 4 | |
| 5 | sys.path.append(os.getcwd()) |
| 6 | |
| 7 | import thop |
| 8 | import torch |
| 9 | |
| 10 | from f5_tts.model import CFM, DiT |
| 11 | |
| 12 | |
| 13 | """ ~155M """ |
| 14 | # transformer = UNetT(dim = 768, depth = 20, heads = 12, ff_mult = 4) |
| 15 | # transformer = UNetT(dim = 768, depth = 20, heads = 12, ff_mult = 4, text_dim = 512, conv_layers = 4) |
| 16 | # transformer = DiT(dim = 768, depth = 18, heads = 12, ff_mult = 2) |
| 17 | # transformer = DiT(dim = 768, depth = 18, heads = 12, ff_mult = 2, text_dim = 512, conv_layers = 4) |
| 18 | # transformer = DiT(dim = 768, depth = 18, heads = 12, ff_mult = 2, text_dim = 512, conv_layers = 4, long_skip_connection = True) |
| 19 | # transformer = MMDiT(dim = 512, depth = 16, heads = 16, ff_mult = 2) |
| 20 | |
| 21 | """ ~335M """ |
| 22 | # FLOPs: 622.1 G, Params: 333.2 M |
| 23 | # transformer = UNetT(dim = 1024, depth = 24, heads = 16, ff_mult = 4) |
| 24 | # FLOPs: 363.4 G, Params: 335.8 M |
| 25 | transformer = DiT(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4) |
| 26 | |
| 27 | |
| 28 | model = CFM(transformer=transformer) |
| 29 | target_sample_rate = 24000 |
| 30 | n_mel_channels = 100 |
| 31 | hop_length = 256 |
| 32 | duration = 20 |
| 33 | frame_length = int(duration * target_sample_rate / hop_length) |
| 34 | text_length = 150 |
| 35 | |
| 36 | flops, params = thop.profile( |
| 37 | model, inputs=(torch.randn(1, frame_length, n_mel_channels), torch.zeros(1, text_length, dtype=torch.long)) |
| 38 | ) |
| 39 | print(f"FLOPs: {flops / 1e9} G") |
| 40 | print(f"Params: {params / 1e6} M") |
| 41 |