#!/usr/bin/env python
# coding: utf-8
YearPredictionMSD

Предсказание года выпуска аудиотрека. Features extracted from the 'timbre' features from The Echo Nest API. Выборка состоит из 90 признаков и 515345 объектов.
# In[31]:


import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
get_ipython().run_line_magic('matplotlib', 'inline')


# In[34]:


column_names = ['year', *('average' + str(i) for i in range(12)), *('covariance' + str(i) for i in range(78))]
data = pd.read_csv("YearPredictionMSD.txt", sep = ",", header=None, names = column_names)


# In[35]:


data.head()

You should respect the following train / test split:
train: first 463,715 examples
test: last 51,630 examples
It avoids the 'producer effect' by making sure no song
from a given artist ends up in both the train and test set.
# In[36]:


train = data.iloc[:463715, :]
test = data.iloc[463715:, :]
len(train), len(test)


# In[37]:


first_year, last_year = 1922, 2011
num_years = last_year-first_year+1


# In[38]:


from torch import nn


# In[112]:


model = nn.Sequential()
model.add_module('l1', nn.Linear(90, num_years))
#model.add_module('activ', nn.ReLU())
model.add_module('smax', nn.Softmax(0))


# In[ ]:





# In[113]:


import torch
opt = torch.optim.Adam(model.parameters(), lr=1e-3)


# In[44]:


#check that answers dffer every time
a = np.random.randint(0, len(train), 5)
b = np.random.randint(0, len(train), 5)
a, b


# In[52]:


X = train.iloc[:, 1:].values
Y = train.iloc[:, 0].values


# In[57]:


list(Y)


# In[60]:


Y = list(Y)
for idx in range(len(Y)):
    year=Y[idx]
    Y[idx] = [int(i+first_year == year) for i in range(num_years)]


# In[74]:


Y = torch.tensor(Y, dtype=torch.float32)


# In[114]:


history = []

batch_size = int(len(train)/50)
for i in range(10):#around a thousand samples
    # sample batch_size random data
    ix = np.random.randint(0, len(train), batch_size)
    x_batch = torch.tensor(X[ix], dtype=torch.float32)
    y_batch = torch.tensor(Y[ix], dtype=torch.float32)
    
    # predict probabilities
    y_predicted = model(x_batch)
    
    #assert y_predicted.dim() == 1, "did you forget to select first column with [:, 0]"
    
    # compute loss, just like before
    loss = torch.mean( (y_predicted - y_batch)**2 )
    
    loss.backward()      # add new gradients
    opt.step()           # change weights
    opt.zero_grad()      # clear gradients
    
    history.append(loss.data.numpy())
    
    if i % 1 == 0:
        print("step #%i | mean loss = %.3f" % (i, np.mean(history[-10:])))


# In[ ]:





# In[ ]: