zomux /
deepy
| @@ 135-149 (lines=15) @@ | ||
| 132 | return sampled_l_t, h_t, a_t, wl_grad |
|
| 133 | ||
| 134 | ||
| 135 | def _output_func(self): |
|
| 136 | self.x = self.x.reshape((28, 28)) |
|
| 137 | first_l, wf_grad = self._first_glimpse_sensor(self.x) |
|
| 138 | ||
| 139 | [l_ts, h_ts, a_ts, wl_grads], _ = theano.scan(fn=self._core_network, |
|
| 140 | outputs_info=[first_l, self.h0, None, None], |
|
| 141 | non_sequences=[self.x], |
|
| 142 | n_steps=5) |
|
| 143 | ||
| 144 | self.positions = l_ts |
|
| 145 | self.last_decision = T.argmax(a_ts[-1]) |
|
| 146 | wl_grad = T.sum(wl_grads, axis=0) / wl_grads.shape[0] |
|
| 147 | self.wl_grad = wl_grad |
|
| 148 | self.wf_grad = wf_grad |
|
| 149 | return a_ts[-1].reshape((1,10)) |
|
| 150 | ||
| 151 | def _setup_functions(self): |
|
| 152 | self._relu = build_activation("tanh") |
|
| @@ 150-161 (lines=12) @@ | ||
| 147 | return sampled_l_t, h_t, a_t, wl_grad |
|
| 148 | ||
| 149 | ||
| 150 | def _output_func(self): |
|
| 151 | self.x = self.x.reshape((28, 28)) |
|
| 152 | [l_ts, h_ts, a_ts, wl_grads], _ = theano.scan(fn=self._core_network, |
|
| 153 | outputs_info=[self.l0, self.h0, None, None], |
|
| 154 | non_sequences=[self.x], |
|
| 155 | n_steps=5) |
|
| 156 | ||
| 157 | self.positions = l_ts |
|
| 158 | self.last_decision = T.argmax(a_ts[-1]) |
|
| 159 | wl_grad = T.sum(wl_grads, axis=0) / wl_grads.shape[0] |
|
| 160 | self.wl_grad = wl_grad |
|
| 161 | return a_ts[-1].reshape((1,10)) |
|
| 162 | ||
| 163 | def _setup_functions(self): |
|
| 164 | self._assistive_params = [] |
|
