7 examples of 'pytorch model summary' in Python

31
def summary(self):
32
    print(self.input_shape)
33
    for l in self.layers:
34
        print(l)
35
        print(l.get_output_shape())

104
def save_summary(self, summary, step, mode):
105
    if mode == 'train':
106
        self.train_writer.add_summary(summary, step)
107
    elif mode == 'valid':
108
        self.valid_writer.add_summary(summary, step)
109
    self.sess.run(tf.local_variables_initializer())

261
def summary(self):
262
    with tf.name_scope('summaries'):
263
        tf.summary.scalar('xent_loss', tf.reduce_sum(self.xent_loss))
264
        tf.summary.scalar('l2_loss', tf.reduce_sum(self.lossL2))
265
        tf.summary.scalar("kl_loss", tf.reduce_sum(self.kl_loss))
266
        tf.summary.scalar("context_kl_loss", tf.reduce_sum(self.context_kl_loss))
267
        tf.summary.scalar('total_loss', tf.reduce_sum(self.cost))            
268
        tf.summary.histogram("latent_vector", self.z_vector)
269
        tf.summary.histogram("latent_mean", self.z_mean)
270
        tf.summary.histogram("latent_log_sigma", self.z_log_sigma)
271
        self.summary_op = tf.summary.merge_all()

415
def __summary(self):
416
    with tf.name_scope('summary'):
417
        self.__mean_accuracy = tf.placeholder(tf.float32, name='mean_accuracy')
418
        self.__mean_loss = tf.placeholder(tf.float32, name='mean_loss')
419
        self.__mean_log_loss = tf.placeholder(tf.float32, name='mean_log_loss')
420
        # self.__mean_ch_log_loss = tf.placeholder(tf.float32, name='mean_ch_log_loss')
421
422
        tf.summary.scalar('learning_rate', self.__learning_rate)
423
        tf.summary.scalar('mean_accuracy', self.__mean_accuracy)
424
        tf.summary.scalar('mean_loss', self.__mean_loss)
425
        tf.summary.scalar('mean_log_loss', self.__mean_log_loss)

103
def summary(self):
104
    x = torch.zeros(BATCH_SIZE, Z_DIM)
105
    
106
    # Print the title in a good design
107
    # for easy recognition.
108
    print()
109
    summary_title = f"| {self.__class__.__name__} Summary |"
110
    for _ in range(len(summary_title)):
111
        print("-", end="")
112
    print()
113
    print(summary_title)
114
    for _ in range(len(summary_title)):
115
        print("-", end="")
116
    print("\n")
117
    
118
    # Run forward pass while not tracking history on
119
    # tape using `torch.no_grad()` for printing the
120
    # output shape of each neural layer operation.
121
    print(f"Input: {x.size()}")
122
    with torch.no_grad():
123
        for layer in self.main1:
124
            x = layer(x)
125
            print(f"Out: {x.size()} \tLayer: {layer}")
126
        x = x.view(-1, 4096, 4, 4)  # Reshape for convolution
127
        print(f"Out: {x.size()} \tLayer: Reshape")
128
        for layer in self.main2:
129
            x = layer(x)
130
            print(f"Out: {x.size()} \tLayer: {layer}")

64
def summary(self):
65
    self.adversarial.summary()
66
    self.discriminator.summary()

153
def log_to_tensorboard(model, step, input_var, losses, top1, topk, mode):
154
    im = {'amount': 10, 'every': 3}
155
    
156
    if step % args.tensorboard_freq == 0 or mode == 'val':
157
        # (1) Log the scalar values
158
        info = {
159
            '{}_loss_avg'.format(mode): losses.avg,
160
            '{}_loss'.format(mode): losses.val,
161
162
            '{}_accr_avg'.format(mode): top1.avg,
163
            '{}_accr'.format(mode): top1.val,
164
165
            '{}_accr_top{}_avg'.format(mode, args.top_k): topk.avg,
166
            '{}_accr_top{}'.format(mode, args.top_k): topk.val
167
        }
168
        for tag, value in info.items():
169
            logger.scalar_summary(tag, value, step)
170
171
        # (2) Log values and gradients of the parameters (histogram)
172
        for tag, value in model.named_parameters():
173
            tag = tag.replace('.', '/')
174
            logger.histo_summary(tag, to_np(value), step)
175
176
            # At the beginning it may happen
177
            # source: https://discuss.pytorch.org/t/zero-grad-optimizer-or-net/1887/6
178
            if value.grad is not None:
179
                logger.histo_summary(tag + '/grad', to_np(value.grad), step)
180
181
182
        # (3) Log the images
183
        # Take non-repeating images (every=3) in specified amount
184
        images = input_var.view(-1, args.image_size, args.image_size)[:(im['amount'] * im['every'])]
185
        images = images[::im['every']]
186
        info = {
187
            '{}_images'.format(mode): to_np(images)
188
        }
189
        for tag, images in info.items():
190
            logger.image_summary('{}_{}'.format(mode, tag), images, step)