File size: 7,158 Bytes
65ffd92
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os, torch, torch.nn as nn, torch.utils.data as data, torchvision as tv
import lightning as L
import numpy as np, pandas as pd, matplotlib.pyplot as plt
from pytorch_lightning.loggers import WandbLogger
import wandb
import pytorch_lightning as pl

torch.set_float32_matmul_precision('medium')

# create the datasets and dataloaders
train_voxels_path = '/fsx/proj-fmri/ckadirt/b2m/data/sub-001_Resp_Training.npy' # path to training voxels 65000 * 4800 
test_voxels_path = '/fsx/proj-fmri/ckadirt/b2m/data/sub-001_Resp_Test_Mean.npy' # path to test voxels 65000 * 600

train_embeddings_path = '/fsx/proj-fmri/ckadirt/b2m/data/encodec_training_embeds_sorted.npy' # path to training embeddings 480 * 2 * 1125
test_embeddings_path = '/fsx/proj-fmri/ckadirt/b2m/data/encodec_testing_embeds_sorted.npy' # path to test embeddings 600 * 2 * 1125

class VoxelsDataset(data.Dataset):
    def __init__(self, voxels_path, embeddings_path):
        # transpose the two dimensions of the voxels data to match the embeddings data
        self.voxels = torch.from_numpy(np.load(voxels_path)).float().transpose(0, 1)
        self.embeddings = torch.from_numpy(np.load(embeddings_path))
        # as each stimulus has been exposed for 15 seconds and the fMRI data is sampled every 1.5 seconds, we take 10 samples per stimulus
        self.len = len(self.voxels) // 10
        print("The len is ", self.len  )

    def __getitem__(self, index):
        # as each stimulus has been exposed for 15 seconds and the fMRI data is sampled every 1.5 seconds, we take 10 samples per stimulus
        voxels = self.voxels[index*10:(index+1)*10]
        embeddings = self.embeddings[index]
        return voxels, embeddings

    def __len__(self):
        return self.len
    
class VoxelsEmbeddinsEncodecDataModule(pl.LightningDataModule):
    def __init__(self, train_voxels_path, train_embeddings_path, test_voxels_path, test_embeddings_path, batch_size=8):
        super().__init__()
        self.train_voxels_path = train_voxels_path
        self.train_embeddings_path = train_embeddings_path
        self.test_voxels_path = test_voxels_path
        self.test_embeddings_path = test_embeddings_path
        self.batch_size = batch_size

    def setup(self, stage=None):
        self.train_dataset = VoxelsDataset(self.train_voxels_path, self.train_embeddings_path)
        self.test_dataset = VoxelsDataset(self.test_voxels_path, self.test_embeddings_path)

    def train_dataloader(self):
        return data.DataLoader(self.train_dataset, batch_size=self.batch_size, shuffle=True)

    def val_dataloader(self):
        return data.DataLoader(self.test_dataset, batch_size=self.batch_size, shuffle=False)


class MLP(pl.LightningModule):
    def __init__(self, sizes, residual_conections, dropout):
        # sizes is a list of the sizes of the layers ej: [4800, 1000, 1000, 1000, 1000, 1000, 1000, 600]
        # residual_conections is a list with the same length as sizes, each element is a list of the indexes of the layers that will recieve the output of the layer as input, 0 means that the layer will recieve the x inputs ej. [[0], [1], [2,1], [3], [4,3], [5], [6,5], [7]]
        # dropout is a list with the same length as sizes, each element is the dropout probability of the layer ej. [0.5, 0.5, 0.5, 0.5, 0.5, 0.5]
        super().__init__()
        self.sizes = sizes
        self.residual_conections = residual_conections
        self.dropout = dropout
        self.layers = nn.Sequential()
        for i in range(len(sizes)-1):
            self.layers.add_module('linear'+str(i), nn.Linear(sizes[i], sizes[i+1]))
            self.layers.add_module('relu'+str(i), nn.ReLU())
            self.layers.add_module('dropout'+str(i), nn.Dropout(dropout[i]))

        self.loss = nn.CrossEntropyLoss(reduction='mean')

    def forward(self, x):
        return self.layers(x)
    
    def training_step(self, batch, batch_idx):
        voxels, embeddings = batch # the sizes are [batch_size, 10, 65000] and [batch_size, 2, 1125]
        # flatten the voxels to [batch_size, rest of the dimensions]
        embeddings = embeddings.flatten(start_dim=1).long() # the size is [batch_size, 2250] 
        #take just the first 200 embeddings
        # embeddings = embeddings[:, :200]
        # take the mean of the second dimension of the voxels to get the mean of the 10 samples per stimulus
        voxels = voxels.mean(dim=1)
        voxels = voxels.flatten(start_dim=1) # the size is [batch_size, 65000]
        outputs = self(voxels)
        # the outputs are [batch_size, 200*1024], we need to reshape them to [batch_size, 200, 1024]
        outputs = outputs.reshape(-1, 1024, 1125*2)
        # avoid division by zero
        outputs = outputs + 1e-6
        #print(outputs.shape, embeddings.shape)
        #print(outputs[0,0,:10], embeddings[0,:10])
        loss = self.loss(outputs, embeddings)
        #print(loss)
        acuracy = self.tokens_accuracy(outputs, embeddings)
        self.log('train_loss', loss)
        self.log('train_accuracy', acuracy)
        return loss
    
    def tokens_accuracy(self, outputs, embeddings):
        # outputs is [batch_size, 1024, 200]
        # embeddings is [batch_size, 200]
        # we need to get the index of the maximum value of each token
        outputs = outputs.argmax(dim=1)
        # now we need to compare the outputs with the embeddings
        return (outputs == embeddings).float().mean()

    
    def validation_step(self, batch, batch_idx):
        voxels, embeddings = batch
        embeddings = embeddings.flatten(start_dim=1).long()
        #embeddings = embeddings[:, :200]
        voxels = voxels.mean(dim=1)
        voxels = voxels.flatten(start_dim=1)
        outputs = self(voxels)
        outputs = outputs.reshape(-1, 1024, 1125*2)
        loss = self.loss(outputs, embeddings)
        accuracy = self.tokens_accuracy(outputs, embeddings)
        self.log('val_loss', loss)
        self.log('val_accuracy', accuracy)
        return loss
    
        
    def configure_optimizers(self):
        return torch.optim.Adam(self.trainer.model.parameters(), lr=2e-5, weight_decay=3e-3)
    

# create the model
sizes = [60784, 1000, 1000, 1125*2*1024]
residual_conections = [[0], [1], [2], [3]]
dropout = [0.5, 0.5, 0.5, 0.5]
model = MLP(sizes, residual_conections, dropout)


# create the data module
data_module = VoxelsEmbeddinsEncodecDataModule(train_voxels_path, train_embeddings_path, test_voxels_path, test_embeddings_path, batch_size=4)

wandb.finish()
from pytorch_lightning.strategies import DeepSpeedStrategy

wandb_logger = WandbLogger(project='brain2music', entity='ckadirt')

# define the trainer
trainer = pl.Trainer(accelerator="gpu", devices = [0,1,2,3,4,5,6,7], max_epochs=1000, logger=wandb_logger, precision='32', strategy=DeepSpeedStrategy(stage=3, logging_batch_size_per_gpu=8), enable_checkpointing=False, log_every_n_steps=10)
#trainer = pl.Trainer(accelerator="gpu", devices = [0,1,2,3], max_epochs=1000, logger=wandb_logger, precision='bf16', strategy='fsdp', enable_checkpointing=False, log_every_n_steps=10)
# train the model
trainer.fit(model, datamodule=data_module)