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

# # Genre recognition: experiment

# Goal: Check the influence of the distance metric (euclidean or cosine).
# 
# Conclusion: The cosine metric seems more appropriate.
# 
# Observations:
# * Higher accuracy with the cosine distance.
# * Why are the atoms and the accuracy different from the experiment on $\lambda_d$ ? With the initialized seed the results should really be the same. The baseline is however strictly the same.
# * Ran for 2h10.

# ## Hyper-parameters

# ### Parameter under test

# In[1]:


Pname = 'dm'
Pvalues = ['cosine', 'euclidean']

# Regenerate the graph or the features at each iteration.
regen_graph = True
regen_features = True


# ### Model parameters

# In[2]:


p = {}

# Preprocessing.

# Graph.
p['K'] = 10 + 1  # 5 to 10 + 1 for self-reference
p['dm'] = 'cosine'
p['Csigma'] = 1
p['diag'] = True
p['laplacian'] = 'normalized'

# Feature extraction.
p['m'] = 128  # 64, 128, 512
p['ls'] = 1
p['ld'] = 10
p['le'] = None
p['lg'] = 100

# Classification.
p['scale'] = None
p['Nvectors'] = 6
p['svm_type'] = 'C'
p['kernel'] = 'linear'
p['C'] = 1
p['nu'] = 0.5


# ### Numerical parameters

# In[3]:


# HDF5 data stores.
p['folder'] = 'data'
p['filename_gtzan'] = 'gtzan.hdf5'
p['filename_audio'] = 'audio.hdf5'
p['filename_graph'] = 'graph.hdf5'
p['filename_features'] = 'features.hdf5'

# Dataset (10,100,644 | 5,100,149 | 2,10,644).
p['Ngenres'] = 5
p['Nclips'] = 100
p['Nframes'] = 149

# Graph.
p['tol'] = 1e-5

# Feature extraction.
p['rtol'] = 1e-5  # 1e-3, 1e-5, 1e-7
p['N_inner'] = 500
p['N_outer'] = 50

# Classification.
p['Nfolds'] = 10
p['Ncv'] = 40
p['dataset_classification'] = 'Z'


# ## Processing

# In[4]:


import numpy as np
import time

texperiment = time.time()

# Result dictionary.
res = ['accuracy', 'accuracy_std']
res += ['sparsity', 'atoms']
res += ['objective_g', 'objective_h', 'objective_i', 'objective_j']
res += ['time_features', 'iterations_inner', 'iterations_outer']
res = dict.fromkeys(res)
for key in res.keys():
    res[key] = []

def separator(name, parameter=False):
    if parameter:
        name += ', {} = {}'.format(Pname, p[Pname])
    dashes = 20 * '-'
    print('\n {} {} {} \n'.format(dashes, name, dashes))
    # Fair comparison when tuning parameters.
    # Randomnesses: dictionary initialization, training and testing sets.
    np.random.seed(1)


# In[5]:


#%run gtzan.ipynb
#%run audio_preprocessing.ipynb
if not regen_graph:
    separator('Graph')
    get_ipython().run_line_magic('run', 'audio_graph.ipynb')
if not regen_features:
    separator('Features')
    get_ipython().run_line_magic('run', 'audio_features.ipynb')

# Hyper-parameter under test.
for p[Pname] in Pvalues:

    if regen_graph:
        separator('Graph', True)
        get_ipython().run_line_magic('run', 'audio_graph.ipynb')
    if regen_features:
        separator('Features', True)
        p['filename_features'] = 'features_{}_{}.hdf5'.format(Pname, p[Pname])
        get_ipython().run_line_magic('run', 'audio_features.ipynb')
    separator('Classification', True)
    get_ipython().run_line_magic('run', 'audio_classification.ipynb')
    
    # Collect results.
    for key in res:
        res[key].append(globals()[key])

# Baseline, i.e. classification with spectrograms.
p['dataset_classification'] = 'X'
p['scale'] = 'minmax'  # Todo: should be done in pre-processing.
if not regen_graph and not regen_features:
    # Classifier parameters are being tested.
    for p[Pname] in Pvalues:
        separator('Baseline', True)
        get_ipython().run_line_magic('run', 'audio_classification.ipynb')
else:
    separator('Baseline')
    get_ipython().run_line_magic('run', 'audio_classification.ipynb')
res['baseline'] = len(Pvalues) * [accuracy]
res['baseline_std'] = accuracy_std


# ## Results

# In[6]:


print('{} = {}'.format(Pname, Pvalues))
for key, value in res.items():
    if key is not 'atoms':
        print('res[\'{}\'] = {}'.format(key, value))

def plot(*args, **kwargs):
    plt.figure(figsize=(8,5))
    x = range(len(Pvalues))
    log = 'log' in kwargs and kwargs['log'] is True
    pltfunc = plt.semilogy if log else plt.plot
    params = {}
    params['linestyle'] = '-'
    params['marker'] = '.'
    params['markersize'] = 10
    for i, var in enumerate(args):
        if 'err' in kwargs:
            pltfunc = plt.errorbar
            params['yerr'] = res[kwargs['err'][i]]
            params['capsize'] = 5
        pltfunc(x, res[var], label=var, **params)
        for i,j in zip(x, res[var]):
            plt.annotate('{:.2f}'.format(j), xy=(i,j), xytext=(5,5), textcoords='offset points')
    margin = 0.25 / (len(Pvalues)-1)
    params['markersize'] = 10
    plt.xlim(-margin, len(Pvalues)-1+margin)
    plt.title('{} vs {}'.format(', '.join(args), Pname))
    plt.xlabel(Pname)
    plt.ylabel(' ,'.join(args))
    plt.xticks(x, Pvalues)
    plt.grid(True); plt.legend(loc='best'); plt.show()

# Classification results.
res['chance'] = len(Pvalues) * [100./p['Ngenres']]
res['chance_std'] = 0
err=['accuracy_std', 'baseline_std', 'chance_std']
plot('accuracy', 'baseline', 'chance', err=err)

# Features extraction results.
if regen_features:
    plot('objective_g', 'objective_i', 'objective_j', log=True)
    # Unweighted objectives.
    print('g(Z) = ||X-DZ||_2^2, h(Z) = ||Z-EX||_2^2, i(Z) = ||Z||_1, j(Z) = tr(Z^TLZ)')
    div = np.array(Pvalues if Pname is 'ld' else p['ld'])
    res['objective_g_un'] = res['objective_g'] / div
    div = np.array(Pvalues if Pname is 'ls' else p['ls'])
    res['objective_i_un'] = res['objective_i'] / div
    div = np.array(Pvalues if Pname is 'lg' else p['lg'])
    res['objective_j_un'] = res['objective_j'] / div
    plot('objective_g_un', 'objective_i_un', 'objective_j_un', log=True)
    plot('sparsity')
    plot('time_features')
    plot('iterations_inner')
    plot('iterations_outer')

    for i, fig in enumerate(res['atoms']):
        print('Dictionary atoms for {} = {}'.format(Pname, Pvalues[i]))
        fig.show()

print('Experiment time: {:.0f} seconds'.format(time.time() - texperiment))