Fix a few more issues related to #216 w/ TResNet (space2depth) and FP16 weights in wide resnets. Also don't completely dump pretrained weights in in_chans != 1 or 3 cases.

Author: rwightman

tests/test_models.py

@@ -115,8 +115,9 @@ def test_model_default_cfgs(model_name, batch_size):
     @pytest.mark.parametrize('model_name', list_models(pretrained=True))
     @pytest.mark.parametrize('batch_size', [1])
     def test_model_load_pretrained(model_name, batch_size):
-        """Run a single forward pass with each model"""
-        create_model(model_name, pretrained=True)
+        """Create that pretrained weights load, verify support for in_chans != 3 while doing so."""
+        in_chans = 3 if 'pruned' in model_name else 1  # pruning not currently supported with in_chans change
+        create_model(model_name, pretrained=True, in_chans=in_chans)
 
 
 EXCLUDE_JIT_FILTERS = [

timm/models/helpers.py

@@ -1,5 +1,10 @@
+""" Model creation / weight loading / state_dict helpers
+
+Hacked together by / Copyright 2020 Ross Wightman
+"""
 import logging
 import os
+import math
 from collections import OrderedDict
 from copy import deepcopy
 from typing import Callable
@@ -86,11 +91,40 @@ def load_pretrained(model, cfg=None, num_classes=1000, in_chans=3, filter_fn=Non
 
     if in_chans == 1:
         conv1_name = cfg['first_conv']
-        _logger.info('Converting first conv (%s) from 3 to 1 channel' % conv1_name)
+        _logger.info('Converting first conv (%s) pretrained weights from 3 to 1 channel' % conv1_name)
         conv1_weight = state_dict[conv1_name + '.weight']
-        state_dict[conv1_name + '.weight'] = conv1_weight.sum(dim=1, keepdim=True)
+        # Some weights are in torch.half, ensure it's float for sum on CPU
+        conv1_type = conv1_weight.dtype
+        conv1_weight = conv1_weight.float()
+        O, I, J, K = conv1_weight.shape
+        if I > 3:
+            assert conv1_weight.shape[1] % 3 == 0
+            # For models with space2depth stems
+            conv1_weight = conv1_weight.reshape(O, I // 3, 3, J, K)
+            conv1_weight = conv1_weight.sum(dim=2, keepdim=False)
+        else:
+            conv1_weight = conv1_weight.sum(dim=1, keepdim=True)
+        conv1_weight = conv1_weight.to(conv1_type)
+        state_dict[conv1_name + '.weight'] = conv1_weight
     elif in_chans != 3:
-        assert False, "Invalid in_chans for pretrained weights"
+        conv1_name = cfg['first_conv']
+        conv1_weight = state_dict[conv1_name + '.weight']
+        conv1_type = conv1_weight.dtype
+        conv1_weight = conv1_weight.float()
+        O, I, J, K = conv1_weight.shape
+        if I != 3:
+            _logger.warning('Deleting first conv (%s) from pretrained weights.' % conv1_name)
+            del state_dict[conv1_name + '.weight']
+            strict = False
+        else:
+            # NOTE this strategy should be better than random init, but there could be other combinations of
+            # the original RGB input layer weights that'd work better for specific cases.
+            _logger.info('Repeating first conv (%s) weights in channel dim.' % conv1_name)
+            repeat = int(math.ceil(in_chans / 3))
+            conv1_weight = conv1_weight.repeat(1, repeat, 1, 1)[:, :in_chans, :, :]
+            conv1_weight *= (3 / float(in_chans))
+            conv1_weight = conv1_weight.to(conv1_type)
+            state_dict[conv1_name + '.weight'] = conv1_weight
 
     classifier_name = cfg['classifier']
     if num_classes == 1000 and cfg['num_classes'] == 1001: