shark/train.py

import torch
from torch import nn
from torch import nn, optim
from torch.utils.data import DataLoader
import numpy as np
import random
import math

import shark

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.lstm = nn.LSTM(
            input_size=8,
            hidden_size=16,
            num_layers=3,
            dropout=0.1,
        )
        self.fc = nn.Linear(16, 1)
        self.out = nn.Sigmoid()

    def forward(self, x, prev_state):
        output, state = self.lstm(x, prev_state)
        logits = self.fc(output)
        val = self.out(logits)
        #print(str(logits.item())+" > "+str(val.item()))
        return val, state

    def init_state(self, sequence_length):
        return (torch.zeros(3, 1, 16),
                torch.zeros(3, 1, 16))

def train(model, seq_len=16*64):
    tid = str(int(random.random()*99999)).zfill(5)
    print("[i] I am "+str(tid))
    ltLoss = 50
    lltLoss = 52
    model.train()

    criterion = nn.BCELoss()
    optimizer = optim.Adam(model.parameters(), lr=0.0001)

    state_h = [None,None]
    state_c = [None,None]
    blob = [None,None]
    correct = [None,None]

    for epoch in range(1024):
        state_h[0], state_c[0] = model.init_state(seq_len)
        state_h[1], state_c[1] = model.init_state(seq_len)

        blob[0], _ = shark.getSample(min(seq_len, 16*(epoch+1)), 0)
        blob[1], _ = shark.getSample(min(seq_len, 16*(epoch+1)), 1)
        optimizer.zero_grad()
        for i in range(len(blob[0])):
            for t in range(2):
                x = torch.tensor([[[float(d) for d in bin(blob[t][i])[2:].zfill(8)]]], dtype=torch.float32)
                y_pred, (state_h[t], state_c[t]) = model(x, (state_h[t], state_c[t]))
                loss = criterion(y_pred[0][0][0], torch.tensor(t, dtype=torch.float32))

                state_h[t] = state_h[t].detach()
                state_c[t] = state_c[t].detach()

                loss.backward()
                optimizer.step()

                correct[t] = round(y_pred.item()) == t
            ltLoss = ltLoss*0.9 + 0.1*loss.item()
            lltLoss = lltLoss*0.9 + 0.1*ltLoss
        print({ 'epoch': epoch, 'loss': loss.item(), 'ltLoss': ltLoss, 'ok0': correct[0], 'ok1': correct[1], 'succ': correct[0] and correct[1], 'acc': str(int(max(0, 1-math.sqrt(lltLoss))*100))+"%" })
        if epoch % 8 == 0:
            torch.save(model.state_dict(), 'model_savepoints/'+tid+'_'+str(epoch)+'.n')

model = Model()

train(model)