zomux /
deepy
| Conditions | 25 |
| Total Lines | 100 |
| Lines | 0 |
| Ratio | 0 % |
| Changes | 1 | ||
| Bugs | 0 | Features | 0 |
Small methods make your code easier to understand, in particular if combined with a good name. Besides, if your method is small, finding a good name is usually much easier.
For example, if you find yourself adding comments to a method's body, this is usually a good sign to extract the commented part to a new method, and use the comment as a starting point when coming up with a good name for this new method.
Commonly applied refactorings include:
If many parameters/temporary variables are present:
Complex classes like FirstGlimpseTrainer.train_func() often do a lot of different things. To break such a class down, we need to identify a cohesive component within that class. A common approach to find such a component is to look for fields/methods that share the same prefixes, or suffixes.
Once you have determined the fields that belong together, you can apply the Extract Class refactoring. If the component makes sense as a sub-class, Extract Subclass is also a candidate, and is often faster.
| 1 | #!/usr/bin/env python |
||
| 67 | def train_func(self, train_set): |
||
| 68 | cost_sum = 0.0 |
||
| 69 | batch_cost = 0.0 |
||
| 70 | counter = 0 |
||
| 71 | total = 0 |
||
| 72 | total_reward = 0 |
||
| 73 | batch_reward = 0 |
||
| 74 | total_position_value = 0 |
||
| 75 | pena_count = 0 |
||
| 76 | for d in train_set: |
||
| 77 | pairs = self.grad_func(*d) |
||
| 78 | cost = pairs[0] |
||
| 79 | if cost > 10 or np.isnan(cost): |
||
| 80 | sys.stdout.write("X") |
||
| 81 | sys.stdout.flush() |
||
| 82 | continue |
||
| 83 | batch_cost += cost |
||
| 84 | |||
| 85 | wl_grad = pairs[1] |
||
| 86 | wf_grad = pairs[2] |
||
| 87 | max_position_value = np.max(np.absolute(pairs[3])) |
||
| 88 | total_position_value += max_position_value |
||
| 89 | last_decision = pairs[4] |
||
| 90 | target_decision = d[1][0] |
||
| 91 | # Compute reward |
||
| 92 | reward = 0.005 if last_decision == target_decision else 0 |
||
| 93 | if max_position_value > 1.8: |
||
| 94 | reward = 0 |
||
| 95 | # if cost > 5: |
||
| 96 | # cost = 5 |
||
| 97 | # reward += (5 - cost) / 100 |
||
| 98 | total_reward += reward |
||
| 99 | batch_reward += reward |
||
| 100 | if self.last_average_reward == 999 and total > 2000: |
||
| 101 | self.last_average_reward = total_reward / total |
||
| 102 | |||
| 103 | if not self.disable_reinforce: |
||
| 104 | self.batch_wl_grad += wl_grad * - (reward - self.last_average_reward) |
||
| 105 | self.batch_wf_grad += wf_grad * - (reward - self.last_average_reward) |
||
| 106 | if not self.disable_backprop: |
||
| 107 | for grad_cache, grad in zip(self.batch_grad, pairs[5:]): |
||
| 108 | grad_cache += grad |
||
| 109 | counter += 1 |
||
| 110 | total += 1 |
||
| 111 | if counter >= self.batch_size: |
||
| 112 | if total == counter: counter -= 1 |
||
| 113 | self.update_parameters(self.last_average_reward < 999) |
||
| 114 | |||
| 115 | # Clean batch gradients |
||
| 116 | if not self.disable_reinforce: |
||
| 117 | self.batch_wl_grad *= 0 |
||
| 118 | self.batch_wf_grad *= 0 |
||
| 119 | if not self.disable_backprop: |
||
| 120 | for grad_cache in self.batch_grad: |
||
| 121 | grad_cache *= 0 |
||
| 122 | |||
| 123 | if total % 1000 == 0: |
||
| 124 | sys.stdout.write(".") |
||
| 125 | sys.stdout.flush() |
||
| 126 | |||
| 127 | # Cov |
||
| 128 | if not self.disable_reinforce: |
||
| 129 | cov_changed = False |
||
| 130 | if batch_reward / self.batch_size < 0.001: |
||
| 131 | if not self.large_cov_mode: |
||
| 132 | if pena_count > 20: |
||
| 133 | self.layer.cov.set_value(self.layer.large_cov) |
||
| 134 | print "[LCOV]", |
||
| 135 | cov_changed = True |
||
| 136 | else: |
||
| 137 | pena_count += 1 |
||
| 138 | else: |
||
| 139 | pena_count = 0 |
||
| 140 | else: |
||
| 141 | if self.large_cov_mode: |
||
| 142 | if pena_count > 20: |
||
| 143 | self.layer.cov.set_value(self.layer.small_cov) |
||
| 144 | print "[SCOV]", |
||
| 145 | cov_changed = True |
||
| 146 | else: |
||
| 147 | pena_count += 1 |
||
| 148 | else: |
||
| 149 | pena_count = 0 |
||
| 150 | if cov_changed: |
||
| 151 | self.large_cov_mode = not self.large_cov_mode |
||
| 152 | self.layer.cov_inv_var.set_value(np.array(LA.inv(self.layer.cov.get_value()), dtype=FLOATX)) |
||
| 153 | self.layer.cov_det_var.set_value(LA.det(self.layer.cov.get_value())) |
||
| 154 | |||
| 155 | # Clean batch cost |
||
| 156 | counter = 0 |
||
| 157 | cost_sum += batch_cost |
||
| 158 | batch_cost = 0.0 |
||
| 159 | batch_reward = 0 |
||
| 160 | if total == 0: |
||
| 161 | return "COST OVERFLOW" |
||
| 162 | |||
| 163 | sys.stdout.write("\n") |
||
| 164 | self.last_average_reward = (total_reward / total) |
||
| 165 | self.turn += 1 |
||
| 166 | return "J: %.2f, Avg R: %.4f, Avg P: %.2f" % ((cost_sum / total), self.last_average_reward, (total_position_value / total)) |
||
| 167 | |||
| 168 |
